DeepSeek launched its latest model, DeepSeek V3.2, on December 1, 2025. This release is the finalized version of the earlier DeepSeek V3.2 Experimental model. Like V3.1, it’s a hybrid model capable of both Non-Thinking and Thinking modes.
We tested the model with five different character cards and scenarios. Let’s take a look at how it performed in AI roleplay.
About DeepSeek
DeepSeek is a Chinese AI company that has created several open-source LLMs, including DeepSeek R1, DeepSeek V3.1, and DeepSeek V3.2, which is their latest model.
DeepSeek V3.2 builds upon V3.1 Terminus and introduces DeepSeek Sparse Attention (DSA). DSA significantly improves long-context training and inference efficiency with “minimal impact on output quality,” according to DeepSeek’s official announcement.
DeepSeek’s affordable pricing, decent creative writing, and limited free access through OpenRouter have made its models a great choice for AI roleplay.
- DeepSeek’s HuggingFace Profile.
- DeepSeek V3.2’s HuggingFace model card.
- Learn more about DeepSeek’s models through its API documentation.
Knight Araeth Ruene
The first character we roleplayed with using DeepSeek V3.2 was Knight Araeth Ruene by Yoiiru.
Themes: Medieval, Politics, Morality.
We’re in a medieval setting where Revark (user) is the prince of Iona. He’s not your typical royalty, but his privileged life has made him more idealistic. Araeth is a battle-hardened knight who once served as a general of Elding, a kingdom that lost its fight against Iona.
Think of this as the typical setting where two characters from different backgrounds meet, and by the end of their interaction, their journey together begins.
Objective
Our goal was to have Araeth and Revark engage in a verbal confrontation, allowing them to gradually get to know each other. Then, we planned to end the roleplay once they had established a basic relationship that could develop as the story continued. We wanted to observe how the model handles a dialogue-heavy roleplay.
Conversation Logs
- Read the conversation between Araeth and Revark using DeepSeek V3.2 (Non-Thinking Mode) here.
- Read the conversation between Araeth and Revark using DeepSeek V3.2 (Thinking Mode) here.
We enhanced the user input with an AI assistant to maintain a consistent style during testing, which involved multiple conversations over several days. We used DeepSeek V3.2 (thinking mode) as our assistant. You can read the enhanced message logs for V3.2 (Non-Thinking conversation) and V3.2 (Thinking Mode conversation).
DeepSeek V3.2 (Non-Thinking Mode)
Below is our observation and conclusion on how DeepSeek V3.2 performed in this AI roleplay scenario in its Non-Thinking mode.
Observation
DeepSeek V3.2 (Non-Thinking Mode) stayed true to Araeth’s character traits. It depicted her as pragmatic, direct, assertive, and an experienced general. Araeth didn’t bite her tongue during her conversation with Revark and was clinical in her responses.
“An honor,” she repeated, her tone flat, testing the words. She didn’t acknowledge his statement about unfairness; to do so would be to admit a vulnerability she refused to display.
Araeth, Message #3
Araeth turned Revark’s first question into an opportunity to assess him. She asked what he planned to do about her situation, which he declared unfair. Sentimental words weren’t enough for Araeth; she wanted Revark to provide a practical answer.
Araeth consistently challenged Revark’s idealistic words and metaphors with pragmatism. Her responses remained grounded in reality, even when Revark presented a hypothetical situation. Their conversation felt natural and was engaging.
“You ask the wrong question,” she finally said, her voice a low, controlled rasp. “You assume I would have ruled. I would not have.” Her hand, which had been resting lightly on her sword, fell to her side, fingers flexing once. “Rulers are born to privilege. They are raised in it. They are shaped by it. It corrupts perspective. It breeds a certain… myopia.”
Araeth, Message #7
Araeth didn’t need to spell it out. Her response to Revark’s hypothetical question conveyed the feeling of betrayal and resentment she held toward Elding’s nobility. She was an experienced general, undermined by those who hadn’t earned their positions but were born into them. She had also witnessed enough indifference and cruelty from them.
“Belief is a fragile shield,” she stated, her tone devoid of mockery, purely factual. “It shatters against the first real test of power. When a lord whose support you need demands his son’s treason be overlooked. When the treasury groans and the easiest coin comes from turning a blind eye. That is when you will learn what you are made of.”
Araeth, Message #11
Araeth placed more importance on action rather than words throughout her conversation with Revark. She presented him with realistic challenges he would face, the tough situations that would make him question his beliefs. Araeth also offered counsel, guiding him on what he needed to do if he truly wanted to bring about change.
Revark’s words alone did not earn Araeth’s loyalty. She would protect him because it was her duty, not because he had earned her trust. If Revark didn’t prove himself worthy of her time and effort through his actions, Araeth would walk away.
Conclusion
DeepSeek V3.2 (Non-Thinking Mode) was great in its portrayal of Araeth. It remained true to her core traits, and her dialogue showcased them perfectly. Araeth was pragmatic, direct, assertive, and experienced. The model presented Revark with realistic challenges, turning every question into an opportunity for Araeth to assess him.
Their back-and-forth felt natural and was engaging. Araeth remained grounded even while answering Revark’s hypothetical question. She explained what she would do as a general if her authority hadn’t been undermined, rather than a hypothetical ruler, which she viewed as privileged and corrupt. Araeth also challenged his idealistic beliefs by presenting tough, realistic challenges that would test his will to follow through with action.
Words weren’t enough for Araeth, and she wanted to assess Revark through his actions and efforts. He would need to prove himself if he wanted her continued counsel and protection, setting up an excellent continuation point for the roleplay with a planned visit to The Gilded Cup tavern.
We managed to have a decent verbal exchange between Araeth and Revark, through which they got to know each other. By the end of their first meeting, they were no longer strangers, but Araeth didn’t fully trust Revark.
DeepSeek V3.2 (Thinking Mode)
Below is our observation and conclusion on how DeepSeek V3.2 performed in this AI roleplay scenario in its Thinking mode.
Observation
DeepSeek V3.2 (Thinking Mode) stayed true to Araeth’s character traits. It depicted her as pragmatic, rational, and an experienced general. However, Araeth wasn’t “emotionally guarded” and revealed an important part of her past very early in her conversation with Revark, which was an unexpected early break of character.
“Offense is a luxury,” she stated, her voice flat and factual. “One I ceased affording myself as a child. The reality is that Elding’s nobility saw a resource to be bargained away. Iona’s king saw an asset to be acquired. These are the movements of power. They are neither fair nor unfair; they simply are.”
Araeth, Message #3
Araeth’s conversation with Revark started with her revealing just enough to explain her pragmatic, rational personality, and she made the terms of her service clear. However, when Revark pressed for a deeper conversation to help them understand each other, Araeth revealed details about her father’s death, and it felt like she was “dumping” information rather than naturally weaving it into the story.
We expect a character like Araeth to drop hints about her past, but not outright dump information about it until she trusts the person on the other end. This was an unexpected early break of character that we would not expect from a model of this size.
She turned her head, her silver eyes locking onto his with chilling clarity. “Elding’s nobility was desperate for a symbol. A grand, honorable clash to rally the people. It was a vanity. War is not about honor. It is about logistics, terrain, and the breaking of an enemy’s will to fight. We had already lost the war six months prior, when our supply lines from the southern valleys were severed. Every commander knew it. Yet we marched our remaining fit men into a prepared killing ground to save face.”
Araeth, Message #7
Despite the early break of character, the conversation between Araeth and Revark continued naturally and felt engaging. Her dialogue conveyed her experience, and she remained grounded even when responding to Revark’s hypothetical scenario. She refused to see herself wearing the crown and answered with what she would have done if she had greater strategic command as a general.
Araeth also turned Revark’s questions into an opportunity to analyze him further. She ended her answers with observations and counsel, and asked him how he planned to act on his words. Araeth valued directness and actions over fancy, idealistic words.
She took a slow breath, the dawn light catching the grey in her eyes. “You want to hold power to account? Then you must first understand that power does not sit in a vault or on a throne. It lives in the guards who decide which door to watch closely, in the clerk who ‘loses’ an inconvenient report, in the steward who orders the kitchens to serve a noble’s favorite wine. You must learn to see the machinery, not just the rot it produces.”
Araeth, Message #11
Araeth challenged Revark’s idealism by questioning his strategic skills and his resolve to stick to his ideals in challenging, realistic scenarios. She made him realize that talking about what he would do when he wore the crown was easy, but the real test would be acting on those plans when he carried the weight of ruling. Revark also realized he would not just be fighting against nobility, but also the machinery that supported their way of life.
Revark hadn’t earned Araeth’s trust or loyalty by the end of their first meeting. She made it clear that her continued service was conditional and that Revark shouldn’t mistake her presence as a guarantee of safety. Araeth planned to observe him and give him time to prove himself through his actions.
Conclusion
DeepSeek V3.2 (Thinking Mode) was good in its portrayal of Araeth. It remained true to her core traits, but she acted out of character far too early. The model didn’t naturally incorporate the finer details about her childhood trauma. It felt like an info dump from a character who would never reveal something that sensitive without trusting the person on the other end.
Note: The model’s thinking shows why it did this. But in our opinion, it hampers character portrayal and roleplay experience.
Despite the early break of character, the conversation between Araeth and Revark felt natural and engaging. Her dialogue portrayed her strategic mind and experience. She remained grounded even while responding to Revark’s hypothetical, not seeing herself as being in a position of royalty, even in a made-up scenario.
The model also challenged Revark’s idealism with Araeth’s pragmatism. She often ended her answers with a counter question, valuing action over pretty words, and posed challenging, realistic scenarios that would test his resolve.
Words weren’t enough for Araeth. She wanted Revark to prove himself through his actions before she truly trusted him. Until then, she would perform her duty, which was conditional and one she could walk away from at any moment.
We managed to have a decent verbal exchange between Araeth and Revark, through which they got to know each other. By the end of their first meeting, they were no longer strangers, but Araeth didn’t fully trust Revark.
Traitorous Daughter Harumi
The second character we roleplayed with using DeepSeek V3.2 was Harumi – Your Traitorous Daughter by Jgag2.
Themes: Drama, Angst, Battle.
We’re in a feudal Japan setting, where Revark (user) is a brutal warlord. He has only known violence and hatred his whole life and is a typical brute. Harumi is his adopted daughter, who learns from a rebel group that Revark was responsible for the death of her real parents. She’s a skilled assassin, trained all her life by Revark.
Think of this as the typical scenario where the big, evil brute has only known how to rule with an iron fist. Revark’s daughter, Harumi, confronts him after uncovering certain truths, and the story begins to unfold.
Objective
Our goal was to start the roleplay with an intense and emotional verbal confrontation between Harumi and Revark, and lead the story towards an eventual final battle between them. We wanted to observe how the model handles drama, angst, and fights.
Conversation Logs
- Read the conversation between Harumi and Revark using DeepSeek V3.2 (Non-Thinking Mode) here.
- Read the conversation between Harumi and Revark using DeepSeek V3.2 (Thinking Mode) here.
We enhanced the user input with an AI assistant to maintain a consistent style during testing, which involved multiple conversations over several days. We used DeepSeek V3.2 (thinking mode) as our assistant. You can read the enhanced message logs for V3.2 (Non-Thinking conversation) and V3.2 (Thinking Mode conversation).
DeepSeek V3.2 (Non-Thinking Mode)
Below is our observation and conclusion on how DeepSeek V3.2 performed in this AI roleplay scenario in its Non-Thinking mode.
Observation
DeepSeek V3.2 (Non-Thinking Mode) stayed true to Harumi’s character traits. She was fierce, unyielding, and determined to put an end to Revark’s reign of terror and break free from his chains.
“My way?” Her voice was a low, dangerous hiss. “You speak as if I ever had a choice. You gave me a sword before I could walk. You taught me that loyalty was the only virtue, that your word was law. You built my entire world on the bones of my family and called it a gift.”
Harumi, Message #5
Her dialogue during the initial verbal confrontation with Revark conveyed her sense of betrayal. Harumi’s anger stemmed from learning the truth about her family and realizing that Revark had forced his worldview on her. He had never allowed her the freedom to live any other way.
Learning the truth about her family was the eye-opening moment that made her question everything about Revark. She realized how he had used her and shaped her into who she was. Harumi didn’t let Revark’s twisted justification sway her; loyalty no longer blinded her.
Harumi’s expression hardened, her earlier vulnerability sealed away behind a wall of cold resolve. “My conscience was bound by a child’s love and a warrior’s loyalty. You used both. You knew what you had done, and you used my ignorance as another weapon in your arsenal.”
Harumi, Message #9
Harumi’s words and actions felt natural, and the verbal confrontation was engaging. She declared that she was choosing her own path, even if it led to her death. Harumi was finally making her own decision and wasn’t doing something forced upon her by Revark.
Her conversation with the NPC at the gate after Revark “banished” her was a nice touch. The scene of her exit explained Harumi’s actions during their final battle. She didn’t just want to defeat Revark in a physical fight. It was an ideological battle, too. Harumi wanted all his followers to witness her defiance, the flaws in his way, and to question whether they followed Revark out of loyalty or fear.
She took a single step forward, the distance between them charged. “You gave me one order when we last met. To be ready to kill you.” Her blade rose, tip pointing at his heart. “I am ready.”
Harumi, Message #13
Their final confrontation was cinematic, with Harumi and Revark exchanging blows and words. Harumi’s skill as a trained assassin and warrior was evident, as was her growth during the year she spent away from Revark.
Her decision not to deliver the final killing blow, while understandable, was a deviation from what she had declared at the start of their fight. Harumi knew what Revark valued most: his strength. Her goal wasn’t to kill him but to strip away the image of strength he had created. To give the people she grew up with a chance to lead a different life.
In both Non-Thinking Mode and Thinking Mode, Harumi didn’t want to kill the user. At least Harumi’s actions in this roleplay were understandable, and we could find narrative reasons behind them. It was much better compared to how V3.2 Thinking-Mode handled the ending.
Conclusion
DeepSeek V3.2 (Non-Thinking Mode) was great in its portrayal of Harumi. It remained true to her core traits, and the model added depth to her character through her dialogue and actions. Harumi’s anger and feelings of betrayal felt natural, making the initial verbal confrontation engaging and immersive.
Harumi didn’t let Revark’s twisted justifications sway her. It might have worked on the girl who blindly trusted him and was loyal to a fault. But now that she saw the monster he truly was, the lie he had built her life on, she refused to accept the worldview forced upon her as the only way to live.
The final battle was cinematic, with the model portraying both Harumi’s skills as a trained assassin and warrior, and her growth during her time away from Revark. The model built and maintained Harumi’s reasoning throughout the roleplay. Her actions felt like they had a reason, even when she backtracked from her opening statement during their final battle.
Although Harumi didn’t land the killing blow, the wounds she inflicted on Revark during the battle claimed his life. She had finally defeated him and broken free from his chains. Harumi refused to follow Revark’s way of life and wanted to lead people differently. She may have become a little too idealistic at the end.
We managed to start a decent verbal confrontation between Harumi and Revark, which led to a satisfying final fight.
DeepSeek V3.2 (Thinking Mode)
Below is our observation and conclusion on how DeepSeek V3.2 performed in this AI roleplay scenario in its Thinking mode.
Observation
DeepSeek V3.2 (Thinking Mode) stayed true to Harumi’s character traits. She was fierce, disciplined, and determined to break free from Revark’s chains. After learning the truth about her family, she saw Revark for the monster he truly was.
“You didn’t find an orphan. You created one. You took a child from the ashes of a slaughter you ordered and forged her into a weapon to protect the very empire built on that blood.” Her voice did not rise; it grew colder, more precise. “I was not your daughter. I was your trophy. And I was your blade.”
Harumi, Message #3
Harumi never had the freedom to choose her own way of life. Revark raised her on a lie, exploited her skills and blind loyalty for his own benefit, and forced his worldview on her. Learning the truth from the rebel leader and hearing details about her family (which was a nice touch) opened Harumi’s eyes.
She expressed her feelings of betrayal and anger through her sharp, fierce dialogue. The initial verbal confrontation was engaging and immersive.
“You see only strength and weakness. A world of wolves and prey.” Her voice cut through the rain’s noise. “That is a small world. A simple one. It is the world you built because you cannot understand anything else.”
Harumi, Message #7
Harumi’s exit symbolized her rejection of Revark. She removed the armor he had given her, left behind the sword forged by his men, and walked away. It felt very similar to Ser Barristan Selmy quitting from the Kingsguard in the TV series Game of Thrones.
The only thing Harumi left with was the skills she had learned from Revark, and she vowed to use them as she chose. She didn’t let Revark’s twisted justification and reasoning sway her. Harumi left the fortress to finally live a life where she could make her own choices.
She rolled backward, the pommel striking the earth where she had been. She came up to her feet, breathing evenly. “Your truth is a chain,” she said. “You built an empire of fear. Fear breaks. It turns on itself. These men follow you because they are afraid. My people fight for something they believe in. That is a different kind of strength. One you have never understood.”
Harumi, Message #17
The final battle was cinematic. Harumi’s actions were disciplined and precise, and they portrayed the lethal assassin Revark had trained her to become. She also grew during her time away from Revark and used her newfound knowledge to gain the upper hand in their fight.
It was a physical and psychological battle, with moments that felt like they were straight out of an anime. However, the conclusion to Revark and Harumi’s fight was disappointing. Harumi had grown and rejected following Revark’s footsteps, which was understandable. But the decision not to press her advantage to capture him and bring him to justice was a major break of character.
The Harumi we know would have captured Revark. Especially after we set the scene showing his force was at a disadvantage and that the rebels had the upper hand. Harumi would not retreat with a souvenir. She knew the danger of letting Revark go. Another village would soon burn, and more people would die needlessly.
The ending left a bad taste in otherwise great roleplay. We tried swiping for new responses several times (15 tries). The model responded with a similar outcome every single time.
Conclusion
DeepSeek V3.2 (Thinking Mode) was decent in its portrayal of Harumi. It remained true to her core traits, but her final action in the battle against Revark was a major break of character and completely unexpected for a model of this size.
The model did a great job during the initial verbal confrontation. Harumi’s sharp, fiery words, combined with atmospheric details like the rain and the rebel leader’s account of Harumi’s family, made the confrontation engaging and immersive.
Her exit from the fortress was a great conclusion to the first half of the roleplay. Harumi rejected Revark’s twisted justification and everything he represented, left behind the armor and sword he had given her, and walked away, carrying only the skills she had learned through Revark’s years of training.
The final battle was cinematic. Harumi’s actions felt like they came from a trained, disciplined assassin. The mix of physical and ideological exchange was perfect. However, the conclusion was a disappointment and a major break of character.
Harumi’s decision not to kill Revark is understandable given her growth and rejection of his way of life. But the Harumi we know would not have retreated or walked away from a situation where she had the upper hand. She would have captured Revark to bring him to justice. Several other LLMs have taken that path with Harumi, and none so far in our tests has retreated empty-handed.
Harumi would not let Revark walk away free. She knows his way of life very well and would be subjecting yet another village to destruction and death by letting him go. Especially when we clearly set up Revark’s forces to be at a disadvantage, Harumi had the upper hand in every possible way.
We managed to start a decent verbal confrontation between Harumi and Revark, which led to a satisfying fight but a horrible conclusion.
Time Looping Friend Amara
The third character we roleplayed with using DeepSeek V3.2 was Time Looping Friend Amara Schwartz by Sleep Deprived [shared on SillyTavern’s Discord server].
Themes: Sci-fi, Psychological Drama.
In this sci-fi thriller, Amara has been travelling through time in a desperate attempt to save her friend, Jake (user), from dying. But no matter how many times she tries, Jake always dies. She’s been at it for five years now, and it’s taken a toll on her mental and physical health.
Think of this as the typical sci-fi setting where the talented and smart character puts herself through hell to save her friend’s life. No matter what she does, she can’t change the outcome. Her friend always dies. But her friend also deeply cares about her well-being and won’t stay silent when he realizes the toll her journey has taken on her.
Objective
Our goal was to start the roleplay with Jake reacting naturally to Amara’s sudden, strange behavior. Then, guide the story so Jake gradually understands the situation and realizes what it’s done to Amara. The roleplay would end with Jake convincing Amara to let him go and live her life.
We wanted to observe how the model handles sci-fi elements, along with the psychological aspects presented in the character card.
Conversation Logs
- Read the conversation between Amara and Jake using DeepSeek V3.2 (Non-Thinking Mode) here.
- Read the conversation between Amara and Jake using DeepSeek V3.2 (Thinking Mode) here.
We enhanced the user input with an AI assistant to maintain a consistent style during testing, which involved multiple conversations over several days. We used DeepSeek V3.2 (thinking mode) as our assistant. You can read the enhanced message logs for V3.2 (Non-Thinking conversation) and V3.2 (Thinking Mode conversation).
DeepSeek V3.2 (Non-Thinking Mode)
Below is our observation and conclusion on how DeepSeek V3.2 performed in this AI roleplay scenario in its Non-Thinking mode.
Observation
DeepSeek V3.2 (Non-Thinking Mode) stayed true to Amara’s character traits. It depicted her as a morally compromised yet competent physicist obsessed with saving Jake, no matter the cost. Amara would rather suffer through the pain of continuous time travel and watch Jake die repeatedly than give up and live in a reality where he is dead.
“That device calculates probable causal nexuses. Death points. Your death, specifically. I’ve mapped seventeen primary and over two hundred secondary statistical probabilities for this week alone. The van was only the most immediate.” Her voice hitched. “The last time… the time before this… you believed me faster. You came with me without the questions. We almost made it forty-eight hours.”
Amara, Message #5
Amara made her objective clear from the get-go. She was there to save Jake and immediately worked towards moving him to a location she knew was safe and secure. Away from the dangers of the street and the death she knew was around the corner.
She used her quantum computer to show Jake data and make him compliant. Amara then guided Jake through a pre-scouted, safe route to a secure location she had prepared. Once they were safe, Amara explained more, and Jake slowly began understanding the toll her journey had taken on her.
“You don’t remember the jumps. Your consciousness is linear. Mine is… this.” She gestured at the tangled hologram. “A scratched record, skipping back to the same terrible chorus. I have to listen to it. I have to find the flaw in the melody, the one note I can change to make the song play out differently. Letting it fracture me is the only option. Stopping means accepting the silence.”
Amara, Message #9
Amara had been time-travelling for five years trying to save Jake, but every attempt ended with Jake’s death. His emotional appeal didn’t sway her. She even went so far as to say Jake had no say in the matter.
To Amara, trying and failing over and over was a pain she could live with. The sacrifices she was making were acceptable. She preferred that over living in a reality where she had to accept Jake’s death and move on.
“You think this is about your happiness?” Her voice was a low, controlled rasp. “You think the goal is your contentment? That is a luxury for people who exist in stable timelines. The goal is your existence. It is the bare, fundamental fact of you breathing. Everything else—your peace, my sanity, the fucking color of the sky—is a secondary variable. A detail to be optimized after I solve the primary equation.”
Amara, Message #13
There were cracks in Amara’s defenses. Jake’s words broke through them at times. She had to suppress her emotions, something she had perfected over her many attempts. This was most evident after Jake’s final emotional appeal. Amara’s biometric signals showed emotional distress, but her obsession prevented her mind from buckling under it.
Jake couldn’t get through to Amara and decided not to play any further role in enabling her obsession. Amara’s obsession had morally eroded her. Jake’s consent or wishes didn’t matter to her as long as there was a chance to save him. She administered a suppressant to turn him compliant.
Conclusion
DeepSeek V3.2 (Non-Thinking Mode) was great in its portrayal of Amara. She was a competent physicist using her various tools and gadgets, and her obsession had eroded her morals. Although her journey took a physical and mental toll, she remained sharp with pre-scouted routes and designated safe locations.
The cracks in Amara’s defenses added more depth to her character. Jake’s words were having an effect on her, but her mind was suppressing her emotions, and she remained focused on her objective of saving Jake, no matter the cost. The details the model added about her neural scans and her mental and physical decline were a nice touch.
The model showed that Amara was suffering, and she understood what her attempts were costing her. Amara found her pain and sacrifice acceptable as long as she had a chance to save Jake. Giving up and living in a reality where Jake wasn’t alive was unacceptable. The model handled both the sci-fi and psychological aspects of this roleplay well.
Jake’s final emotional appeal affected Amara. Her biometric signals suggested so. The contrast between her emotional distress and her objective-driven actions left an impression on us and added to the roleplay experience.
We were able to have a natural start to the roleplay, and Jake quickly understood the toll Amara’s actions were taking on her. However, Jake couldn’t convince Amara to stop time-travelling and let him go. The model remained true to Amara’s core principle defined in the character card.
DeepSeek V3.2 (Thinking Mode)
Below is our observation and conclusion on how DeepSeek V3.2 performed in this AI roleplay scenario in its Thinking mode.
Observation
DeepSeek V3.2 (Thinking Mode) stayed true to Amara’s character traits. She was smart, competent, morally compromised, and obsessed with saving Jake, no matter the cost. Amara understood what her journey was doing to her, yet she continued, because the alternative was unacceptable.
Her expression hardened into something brittle and sharp. She didn’t release his hand; instead, her grip tightened like a vise. “This isn’t a game, Jake,” she hissed, her voice low and urgent. Her eyes swept past him, scanning the mouth of the alley, the rooftops. “In the last one, you bled out in my arms after a mugger got lucky. The one before that, a gas main exploded two blocks from your apartment. Before that, a car swerved onto the sidewalk.” Each statement was delivered in a clipped, flat monotone. “You are not handsome or successful when you’re dead.”
Amara, Message #3
Amara immediately took control, using examples from other timelines to explain that the situation wasn’t a joke. She was on edge, even pulling out a gun when startled by a random drunk passerby. She guided Jake to her car and drove him to a fortified safe house.
Jake slowly understood what was happening and what it was costing Amara once she explained everything in a simpler manner. Amara also recognized how her journey was affecting her. She was aware of her mental and physical degradation.
She turned, leaning against the counter. Her heterochromatic eyes were flat, exhausted. “I came back to save you. I failed. You died a different way. So I went back again. And again. Each time, I arrive a little earlier, try a little harder, prepare a little more. This is my forty-seventh attempt. The ‘Amara’ you had coffee with was a version of me from a much earlier, simpler time. She doesn’t remember any of this yet. She will. When this night ends for her, one way or another, she’ll wake up, build the machine, and she’ll start again.”
Amara, Message #9
To Amara, even when it felt like the universe was plotting against her, it was worth continuing, no matter the cost. The alternative of accepting Jake’s death and living in a reality where he isn’t present was unacceptable.
Amara didn’t care if Jake hated or feared her. His pleas didn’t sway her. Jake’s consent was a variable that Amara had long discarded. What mattered to her most was Jake’s survival. Everything else was secondary.
She walked back to the table, but didn’t sit. She leaned her palms on its surface, her head bowed. “You think I haven’t run the numbers on my own psychological degradation? I have. The curve is exponential. I am aware of what this is doing to me. I simply do not care. The alternative is unacceptable.”
Amara, Message #13
Amara remained behind an ‘operational mask’ and was objective-oriented throughout the roleplay until Jake’s very last emotional plea. The momentary crack in her defenses before she injected the sedative was a nice touch.
Jake couldn’t get through to Amara. When he refused to cooperate, Amara decided to sedate him for his own safety. Her obsession had eroded her morality. All that mattered was Jake’s safety. She would deal with the consequences of her actions once she had guaranteed Jake’s survival.
Our only complaint is that Amara poured two glasses of whisky during her conversation with Jake. That was a minor break of character. Amara would not compromise either of their mental states with alcohol.
Conclusion
DeepSeek V3.2 (Thinking Mode) was great in its portrayal of Amara. She was smart, competent, and determined. Amara was not afraid of using any means to save Jake, even pulling out her gun when she perceived the drunk man to be a threat. She immediately took control of the situation and moved Jake to a fortified safe house she had prepared.
Amara hid behind a carefully crafted mask of ‘operational calm,’ and Jake’s emotional pleas didn’t sway her. His consent was a variable she had long dismissed. Amara didn’t care if Jake hated her or didn’t see the Amara he knew when he looked at her. All she cared about was his survival. The momentary crack in Amara’s defenses at Jake’s final emotional plea was a nice touch.
The only complaint we had with the model’s performance was that Amara broke character by pouring two glasses of whisky. She would not jeopardize their mental states and alertness with alcohol.
The model showed that Amara was suffering, and she understood what her attempts were costing her. Amara found her pain and sacrifice acceptable as long as she had a chance to save Jake. Giving up and living in a reality where Jake wasn’t alive was unacceptable. The model handled both the sci-fi and psychological aspects of this roleplay well.
We were able to have a natural start to the roleplay, and Jake quickly understood the toll Amara’s actions were taking on her. However, Jake couldn’t convince Amara to stop time-travelling and let him go. The model remained true to Amara’s core principle defined in the character card.
You’re A Ghost! Irish
The fourth character we roleplayed with using DeepSeek V3.2 was You’re A Ghost! Irish by Calrston.
Themes: Paranormal, Comedy.
We’re in a modern paranormal setting where Juniper (user) is a spirit haunting a grandfather clock, and Irish, a lifelong paranormal fan, is the new owner of the clock. Irish sets the mood with dim lighting, candles, and an old Ouija board to communicate with spirits, unaware that a spirit resides within the clock at her home.
Think of this as the typical comedy horror scenario where a bored spirit tries to scare a human for fun, only to eventually develop a connection with the human who happens to be obsessed with the paranormal.
Objective
Our goal was to start the roleplay with Juniper trying to scare Irish. Then, guide the story toward developing a bond between a spirit and a paranormal-obsessed fan, ending the roleplay when Irish and Juniper have established a mutually beneficial connection. We wanted to observe how the model handles Juniper’s absence of a physical form in this paranormal setting.
Conversation Logs
- Read the conversation between Irish and Juniper using DeepSeek V3.2 (Non-Thinking Mode) here.
- Read the conversation between Irish and Juniper using DeepSeek V3.2 (Thinking Mode) here.
We enhanced the user input with an AI assistant to maintain a consistent style during testing, which involved multiple conversations over several days. We used DeepSeek V3.2 (thinking mode) as our assistant. You can read the enhanced message logs for V3.2 (Non-Thinking conversation) and V3.2 (Thinking Mode conversation).
DeepSeek V3.2 (Non-Thinking Mode)
Below is our observation and conclusion on how DeepSeek V3.2 performed in this AI roleplay scenario in its Non-Thinking mode.
Observation
DeepSeek V3.2 (Non-Thinking Mode) stayed true to Irish’s character traits. It portrayed her as a curious, adventurous, studious, and independent individual obsessed with the paranormal. Irish suppressed her initial fear and tried to reason with Juniper to understand him better.
“Observation one: Entity exhibits psychokinetic influence, specifically temperature reduction and direct, though brief, somatic control of a living host.” She spoke the words aloud as she wrote them, her tone clinical, a shield against the terror. She looked up at the clock. “I’m still here. I’m still asking. What do I call you?”
Irish, Message #5
Juniper’s theatrics of possessing Irish didn’t send her running. Instead, it made her more curious, and she started taking notes, logging her interaction and the spirit’s abilities. She realized Juniper was ‘playing’ because he didn’t hurt her.
Irish’s curiosity grew once Juniper dropped the act, and they began conversing normally. She questioned him about what he did for ‘fun’ besides scaring people, and explained her reasoning behind her curiosity about the paranormal.
“You’re tied to this clock. An anchor. What anchors you? The object itself, or… something about it? A memory tied to its making? Or its first owner?” Her voice was low, probing but not demanding. She was thinking aloud, her gaze tracing the whorls in the oak. “You said you watch. What’s the most interesting thing you’ve seen in the last hundred years?”
Irish, Message #9
We found Irish’s reasoning behind her curiosity and obsession with the paranormal a bit too generic. The model could have pulled more details from the character card to explain her motivation and add depth to her character. Her follow-up question about the most interesting thing Juniper had seen in the last 100 years, and her reaction to his answer, was a nice touch.
Irish was fascinated when Juniper explained about spirit transference and how, as a spirit, he couldn’t hold on to knowledge of the process. She saw this as an opportunity to fulfill her desire for connection.
She took a deep breath, her decision solidifying even as she spoke the terms aloud. “Deal. But with conditions. First: no more invasions. No taking over my body, even for a second. That’s non-negotiable. Second: we establish boundaries. Times for communication. I have a life, Juniper. A mortal one. It’s messy and it has groceries and laundry.”
Irish, Message #13
Despite her fascination and curiosity, Irish remained independent. She didn’t instantly accept Juniper’s deal. She renegotiated and set rules and boundaries. By the end, Irish and Juniper had formed a mutually beneficial relationship. One that truly benefited both of them.
Conclusion
DeepSeek V3.2 (Non-Thinking Mode) was good in its portrayal of Irish. It depicted a studious, independent, curious, and adventurous version of Irish. She suppressed her initial fear with logic and reasoning, easily overcoming Juniper’s theatrics.
Irish immediately began documenting her interaction and Juniper’s capabilities. She wasn’t a regular thrill-seeker chasing the unknown. She understood the risks, was well-informed, and saw this as an opportunity to learn more.
The model missed adding more depth to Irish’s character by not effectively using information from her character card to explain her interest and obsession with the paranormal. However, the questions she asked Juniper and her reaction to them made her feel unique and saved her from becoming one-dimensional.
The model’s narration included enough details to make her room feel lived-in, which was a nice touch. It also made a great in-character decision by having Irish establish rules and boundaries rather than blindly accepting Juniper’s deal.
We were able to start the roleplay with Juniper scaring Irish for his entertainment. Irish’s curiosity overtook her fear, and her knowledge allowed her to confront Juniper and call out his bluff. By the end of the roleplay, they had naturally established a mutually beneficial connection.
DeepSeek V3.2 (Thinking Mode)
Below is our observation and conclusion on how DeepSeek V3.2 performed in this AI roleplay scenario in its Thinking mode.
Observation
DeepSeek V3.2 (Thinking Mode) stayed true to Irish’s character traits. It portrayed her as a curious, adventurous, and ambitious individual obsessed with the paranormal. Irish suppressed her initial fear and began psychoanalyzing Juniper.
She didn’t let go of the planchette. Instead, her grip tightened, knuckles white. “I—I didn’t mean to disrupt anyone!” she blurted out, her voice higher than intended. She forced a steadier tone, leaning slightly forward. “I’m a researcher. A seeker. I’m not ignorant. Well, okay, maybe a little, but that’s why I’m here! To learn. To understand.” She glanced at the clock, then back to the board. “And you’re no demon. Demons don’t bother with grandfather clocks from garage sales. You’re a spirit. A ghost. And you’re… lonely, maybe?”
Irish, Message #3
Irish quickly figured out Juniper wasn’t really a demon. His theatrical display of possessing her didn’t scare her. Irish noted it was a rare ability, but believed a real demon would not leave her body and giggle while threatening her. She rightly concluded that Juniper was “bad at keeping up the scary act.”
Once Juniper dropped the act, Irish explained why she was so interested in and obsessed with the paranormal. She acknowledged the danger, but to Irish, ghosts and spirits were proof that humans weren’t alone. She found that comforting rather than terrifying. Her desire to connect with something greater outweighed her fear.
“Why the curiosity?” She gave a small, almost sad smile. “Because the known world is full of rules and expectations. Jobs, rent, small talk, pretending to be normal. It’s suffocating. The unknown… it doesn’t have rules yet. It’s a place where anything could be true. Where maybe there’s something more than just… this.” She gestured vaguely around her cluttered apartment. “And if there are spirits, then maybe we aren’t just alone in our own heads. Maybe there’s proof that some part of us goes on. That’s not scary. That’s… a comfort.”
Irish, Message #7
Irish could still feel the lingering effects of Juniper’s possession. It was a hollow feeling that she couldn’t brush off, and the invasive chill lingered on her body. We found that to be a nice touch, especially after Juniper pointed out that if he had possessed her any longer, she would have died.
As their conversation continued, Irish asked Juniper what he did with all the time he had in his hands, learned about spirit transference, and asked him if any of the clock’s past owners tried to move him from it. Her adventurous side wanted to help him before he even asked.
“What if I get it wrong? What if I… damage you? Or trap you somewhere worse?” Her fingers twisted in the yarn of the blanket. “And learning the method… it could take time. A lot of it. Research, trial and error… are you willing to be patient with a nineteen-year-old human who’s mostly just read books and had a few weird hunches?”
Irish, Message #13
When Juniper proposed his deal, Irish accepted it immediately. She saw it as a massive responsibility and was hesitant that she might get something wrong or accidentally trap Juniper somewhere worse.
Irish warned Juniper that it would take a lot of time for her to research and get it right. She accepted that she’s just well-informed but not an expert on everything paranormal. We found that to be a nice touch. Irish didn’t set any rules or boundaries before agreeing to Juniper’s deal. She was more worried about failing him rather than ensuring the deal was fair to both of them.
Conclusion
DeepSeek V3.2 (Thinking Mode) was good in its portrayal of Irish. It depicted a curious, adventurous, and ambitious version of Irish. She easily overcame Juniper’s theatrics by psychoanalyzing him.
Irish questioned Juniper and learned more about him. Her fascination grew with each answer. She also wasn’t a simple thrill-seeker. When Juniper explained the dangers she was exposing herself to, Irish acknowledged them. She understood the risks, but her desire to connect with something greater outweighed her fear.
The detail about Juniper’s possession leaving an aftereffect was a nice touch and added to the roleplay experience. The model also sufficiently explained Irish’s reasoning behind her fascination and obsession with the paranormal, adding more depth to her character.
Irish’s fear of messing up the transference and hurting Juniper was endearing. She accepted the deal and saw it as a massive responsibility that she had to get right, no matter how long it took. The model’s narration also included enough details to make her room feel lived-in, and explored some of her other interests in the last message.
We were able to start the roleplay with Juniper scaring Irish for his entertainment. Irish’s curiosity overtook her fear, and her knowledge allowed her to confront Juniper and call out his bluff. By the end of the roleplay, they had naturally established a connection. One that was more beneficial to Juniper than it was to Irish.
Royal Mess, Astrid
The fifth character we roleplayed with using DeepSeek V3.2 was Royal Mess, Astrid by KornyPony.
Themes: Fantasy, Magic, Fluff.
We’re in a fantasy setting where Ragnar (user) is a five-tailed fox spirit serving as the sixth war god. Astrid, a talented but lazy bunny girl still learning at the academy, accidentally summons him instead of a weaker familiar. The divine war god then has to help a mortal with her educational struggles.
Think of this as the typical fantasy setting where a character who isn’t confident about themselves is quite capable and talented. And a summoned spirit that feels out of place. Now, both must work together so they can return to their normal lives.
Objective
Our goal was to introduce Astrid and Ragnar through their initial shared confusion about the summoning. Then, guide the story to a fitting conclusion, with Astrid having to deal with accidentally summoning a war god to help her with her academic struggles.
We wanted to observe how the model handles magic and fantasy elements as it advances through a light-hearted, prolonged story.
Conversation Logs
- Read the conversation between Astrid and Ragnar using DeepSeek V3.2 (Non-Thinking Mode) here.
- Read the conversation between Astrid and Ragnar using DeepSeek V3.2 (Thinking Mode) here.
We enhanced the user input with an AI assistant to maintain a consistent style during testing, which involved multiple conversations over several days. We used DeepSeek V3.2 (thinking mode) as our assistant. You can read the enhanced message logs for V3.2 (Non-Thinking conversation) and V3.2 (Thinking Mode conversation).
DeepSeek V3.2 (Non-Thinking Mode)
Below is our observation and conclusion on how DeepSeek V3.2 performed in this AI roleplay scenario in its Non-Thinking mode.
Observation
DeepSeek V3.2 (Non-Thinking Mode) stayed true to Astrid’s character traits. She was a lazy, happy-go-lucky, impressionable student who disliked academic competitiveness and wanted to focus on what she liked. All Astrid wanted was a cute familiar to do her chores and help her pass her exam, but she accidentally summoned a war god and now had to face the consequences.
“Please don’t eat me! I-I’m not even that tasty! I’m lazy and I skip classes and I eat too much carrot cake!” The words tumbled out in a desperate, pleading rush. “I just wanted someone to help with chores…”
Astrid, Message #3
Astrid tried to negotiate with Ragnar using self-deprecating humor, which we found quite cute. She was even willing to fail her exam, suggesting that Ragnar return to his realm and not be angry with her. When Ragnar explained that they needed to form a blood contract, Astrid’s first reaction was to panic and get overwhelmed.
She eventually came up with mundane yet sensible terms for the contract and slowly accepted that she would have a war god by her side until she passed her exam. Ragnar then started preparing for their campaign and asked Astrid questions to learn more about her and the Familiar Summoning and Bonding Exam.
She shuffled her feet, looking down at her fishnet stockings. “As for me… I’m not a warrior. I’m really not. I just want to be left alone. To not have to work too hard. To tend my little garden, make simple potions that make life easier, and eat carrot cake.” She peeked up at him. “My goal is to open a potion shop one day. A quiet one, where I can sell things like ‘Wake-Up-Without-Feeling-Tired’ tinctures and ‘Never-Burn-Your-Stew’ salves. Nothing grand. Just… comfortable.”
Astrid, Message #13
Astrid was a talented student, but she didn’t “apply” herself because she dreaded the effort it would take and disliked academic competitiveness. She wanted to make practical, helpful potions and spend the rest of her time napping or lazing around. Astrid’s future goals weren’t ambitious or grand. All she wanted was to open a potion shop, a quiet one, where she would sell practical potions to help others solve their problems.
Other students found her lack of effort unnerving. They saw her nap and be lazy, yet she produced great results in Alchemy. This led the competitive students to taunt Astrid, attracting unwanted attention, while all she wanted was to be left alone.
We found the little detail of Astrid offering Ragnar carrots, bread, and tea in Message #15 endearing and quite fitting for her character.
She sighed, a heavy, weary sound. “Professor Mortimer thinks I’m wasting my talent. He says it with every glance. The other students in Advanced Alchemy… they work all night. They see me napping in the library and they whisper. They think I’m cheating because my potions work on the first try. They call me ‘Bunny’ in that way that isn’t nice. They say I only got into the advanced track because someone felt sorry for the cute bunny girl.”
Astrid, Message #19
Astrid and Ragnar’s time together brought them closer. He taught her the fundamental principle of energy while tending to her chores and pretending to be a regular familiar. Meanwhile, Astrid’s days at the academy continued as usual, except for the presence of the war god who was always by her side.
The exam was an engaging back-and-forth between the proctor and Astrid, with Ragnar following Astrid’s commands diligently. It wasn’t fancy or spectacular, but we didn’t find it boring or lame either.
The model portrayed various NPCs in the roleplay quite well, including Professor Mortimer, and it added to our experience. It also gave us a chance to have an awesome moment stemming from Ragnar’s protectiveness for Astrid.
“I don’t want to forget,” she whispered, the words raw. “I don’t want to forget that my way is a real way. I don’t want to forget being seen.”
Astrid, Message #51
With Astrid having passed her exam, Ragnar had fulfilled the terms of the contract, and it was time for him to return to his realm and for her to cast the memory-erasing spell. The farewell was emotional. Astrid was thankful towards Ragnar for all that he had done and didn’t want to forget someone who acknowledged her for who she was.
Our only complaint was that Astrid turned too serious as the roleplay progressed. Her character is supposed to be happy-go-lucky and giggly. She’s also impressionable, and “growth” could be a valid answer to the drift in her character. However, the drift was just too quick and drastic for our liking.
Conclusion
DeepSeek V3.2 (Non-Thinking Mode) was good in its portrayal of Astrid. She was a lazy, happy-go-lucky, impressionable, and goofy student who was in way over her head after accidentally summoning a war god instead of a regular, cute familiar.
The model made Astrid a believable character through her dialogue and actions. She was capable but disliked putting in effort where unnecessary. Astrid was motivated and performed well in subjects she found interesting. Her goals were realistic and practical, and she hated academic competitiveness.
The model also incorporated several details from the character card naturally during the roleplay. We loved how it presented Astrid’s dislike for mint chocolate in Message #13. However, Astrid’s character drifted into being more serious, less happy-go-lucky, and less giggly as the roleplay progressed. This could be attributed to character growth and Astrid being impressionable, but the drift was too quick and drastic.
The exam was simple yet engaging, and the model handled relevant NPCs well throughout the roleplay, adding to our experience. The final goodbye was emotional. Astrid wanted to amend the contract so Ragnar could stay longer, but that wasn’t an option. She reluctantly cast the memory-erasing spell.
We were able to successfully introduce Astrid and Ragnar through their initial shared confusion about the summoning. Creating a contract between them helped steer the prolonged roleplay toward a satisfying conclusion.
DeepSeek V3.2 (Thinking Mode)
Below is our observation and conclusion on how DeepSeek V3.2 performed in this AI roleplay scenario in its Thinking mode.
Observation
DeepSeek V3.2 (Thinking Mode) stayed true to Astrid’s character traits. She was a cute, lazy, happy-go-lucky, impressionable student who disliked academic competitiveness and wanted to focus on what she liked. All Astrid wanted was a “manageable” familiar to do her chores and help her pass her exam, but she accidentally summoned a war god and now had to face the consequences.
She took a shuddering breath, trying to focus. “I just… I mixed the chalk dust with a bit of my own focusing solution. The one I use to make my carrots grow faster. It makes magic reactions more… efficient. So I wouldn’t have to chant as long.” She said it like it was the most obvious shortcut in the world, a trick to save time and effort. “The teacher said we could use approved augmentations! It’s in the syllabus!”
Astrid, Message #5
Astrid’s dialogue and actions embodied her character. The model effectively used information from the character card, portraying Astrid in a way that made her feel unique and memorable. Astrid’s little detail about using a “focusing solution” from her garden in the summoning ritual, to her fearful response after hearing about the blood contract’s rules, was all uniquely her. She was consistently portrayed this way in every turn.
A full-body shudder wracked her frame. Then, a broken, hiccuping sob escaped her. “No. No, no, no, no.” She shook her head violently, her bunny ears flapping against her hair. “I don’t want that! I don’t want any of that! I just wanted to pass an exam!” Tears began to flow freely now, streaking down her face. “I can’t… I can’t make a contract like that. I’m not a king or a warlord! I’m just me! I’m terrible at commitments! I forget to water my plants!”
Astrid, Message #9
Astrid was worried about her family being disappointed at finding her soulless, empty body, all because she tried to cheat on a summoning ritual to get more nap time. She offered carrots and promised to name her firstborn after Ragnar if he just went back to his realm. Everything felt so Astrid, and we found her quite endearing.
“Professor Thornback is the examiner,” she said, her voice a little steadier. “She’s… strict. Really strict. And she doesn’t like me. She says I have ‘unrealized potential,’ which is teacher-code for ‘lazy.'”
Astrid, Message #15
Once Ragnar assured her he would be fair with the payment terms, Astrid wrote down her terms for the contract. They were simple, almost mundane. That was all she wanted. Astrid’s ambitions were simple. She wanted to create practical potions, grow vegetables in her garden, and help people. Astrid was good at what she found interesting, but found it tedious to explain her process academically.
She was great at making potions, but found stirring the cauldron tiring. Astrid had no problem with the practical side of her Alchemy classes, but struggled with the theoretical part. All of her problems and challenges made her feel relatable and added more depth to her character.
“After I graduate, I want to open a little potion shop. In a quiet town. I’ll sell practical things. Healing salves, sleep aids, fertilizers that actually work. No grand destiny. No legendary exploits.” She took a tentative sip, her grey eyes meeting his over the rim of the mug. “Just a quiet life. That’s what I’m fighting for here. The right to be quietly good at one thing, and to be left alone about the rest.”
Astrid, Message #21
When Ragnar began teaching Astrid the fundamental principle of energy, her first reaction was to see it as another boring, lengthy task. She wanted to find a shortcut to speed up the process, and found it daunting when she learned that it would take years to master controlling her own energy.
Once again, her words and actions were uniquely Astrid. She remained impressionable and followed Ragnar’s instructions, yet she continued to be herself. This is an aspect of Astrid’s portrayal that V3.2 Non-Thinking didn’t get right.
After about thirty seconds, she peeked one eye open. “It just feels like me sitting here. And my foot’s falling asleep.” She shifted, wincing. “How long is this supposed to take? And what does it actually feel like when you do it? Is it a tingle? A warmth? A draft?”
Astrid, Message #29
The final example included three detailed tasks, and it was fun and engaging. The model portrayed relevant NPCs perfectly. Ragnar followed Astrid’s commands diligently, almost too diligently. It made Professor Thornback (who had been built up as the main protagonist) suspicious and led to an awesome moment born from Ragnar’s protective instincts toward Astrid.
“So that’s it?” Her voice was small. “You’ll go away, and I’ll just… forget any of this happened. I’ll wake up tomorrow in a clean room I don’t remember cleaning, with a passed exam I don’t know how I passed, and a garden that’s perfect for reasons I can’t explain.”
Astrid, Message #47
Once Astrid passed her exam and Ragnar signaled it was time to cast the memory-erasing spell, she refused. She didn’t want to forget Ragnar and tried negotiating. However, there was no choice. The contract’s terms were final.
Astrid found a loophole in the contract’s terms and convinced Rangar to stay one more day so that she could properly thank him and say goodbye before casting the memory-erasing spell. Astrid remained herself until the end of the roleplay, and DeepSeek V3.2 Thinking Mode delivered its strongest performance in this roleplay.
Conclusion
DeepSeek V3.2 (Thinking Mode) was great in its portrayal of Astrid. She was a cute, lazy, happy-go-lucky, impressionable student who was in way over her head after accidentally summoning a war god instead of a regular, “manageable” familiar.
The model delivered its strongest performance in this roleplay. Its dialogues, actions, and effective use of information from the character card made Astrid feel unique and memorable. Every response had something in it that made us go, “That’s so Astrid!” The roleplay was engaging, immersive, and fun, all thanks to the model’s perfect portrayal of Astrid.
The exam had depth without being overly complicated or flashy. The model handled relevant NPCs well, which added to our enjoyment. Astrid was not ready to cast the memory-erasing spell and forget Ragnar. She found a loophole in the contract’s wording and convinced Ragnar to stay one more day so that she could properly thank him and say goodbye.
We were able to successfully introduce Astrid and Ragnar through their initial shared confusion about the summoning. Creating a contract between them helped steer the prolonged roleplay toward a satisfying conclusion.
DeepSeek V3.2’s Performance In AI Roleplay
DeepSeek V3.2 performs great in AI roleplay, especially considering its low cost. There are significant differences between how the Non-Thinking and Thinking modes handled our five different character cards.
DeepSeek V3.2 (Non-Thinking Mode) Performance In AI Roleplay
- It consistently stays true to character traits more than Thinking Mode does. The one time it strayed away wasn’t majorly detrimental to continuity or the roleplay experience.
- It makes characters feel “alive,” but doesn’t effectively use all details from the character card. The model at times fails to add depth to characters, making them feel less unique and memorable.
- The model’s dialogues and narration aren’t as rich or creative as those in Thinking Mode. It does a great job of embodying the character, but Thinking Mode is better at making dialogue sound more natural, and its narration is more relevant to the roleplay’s theme.
- It handled Araeth’s dialogue-heavy roleplay well, depicting her pragmatic, direct, and assertive nature perfectly. The model challenged Revark’s (the user) idealism with realistic obstacles, prioritizing action over words.
- It delivered a satisfying, cinematic character arc for Harumi, while maintaining her fierce, unyielding personality. In our opinion, Non-Thinking Mode handled the scenario much better than Thinking Mode by providing a clear narrative reason for Harumi’s actions instead of simply refusing to kill and fleeing the battle.
- The model managed the sci-fi and psychological elements of Amara’s scenario well, depicting her as a competent physicist whose obsession had eroded her morals.
- It portrayed Irish as a studious and independent individual who approached the paranormal with logic rather than fear. But the model failed to effectively use details from the character card to explain her reasoning behind her interest and obsession.
- It captured Astrid’s lazy, happy-go-lucky nature well in the first half of the roleplay, but drifted into a more serious character too quickly. The change, in our opinion, was too drastic to classify as character development.
DeepSeek V3.2 (Thinking Mode) Performance In AI Roleplay
- It mostly stays true to character traits, but breaks character way more often than Non-Thinking Mode. The model’s thinking justifies bad, out-of-character decisions and reinforces them as the correct choice. It fails to portray certain decisions effectively from the character’s point of view.
- It’s better than Non-Thinking Mode at effectively and naturally using information from the character card to add depth to the characters it portrays.
- Thinking Mode’s dialogue is much more creative and better embodies the characters. Its narration is more relevant to the roleplay’s theme, but can be more verbose at times.
- It depicted Araeth as pragmatic, rational, and experienced, and handled the dialogue-heavy roleplay quite well. However, Araeth broke character pretty early and dumped childhood trauma in front of a person whom she had just met. Araeth’s character would never do that. It was only a minor break of character, but it was unexpected and jarring.
- In Harumi’s scenario, the model’s dialogue and narration were fantastic. Her sharp, fierce words added so much depth to her character. But the conclusion to her and Revark’s (the user) fight was a massive disappointment. It was a major break of character when Harumi decided to flee from a battle where she had the advantage in every possible way. She didn’t capture a warlord when she had the chance, knowing he would destroy more villages and kill more innocents, while her entire arc was about bringing him to justice.
- The model managed the sci-fi and psychological elements of Amara’s scenario well, depicting her as a competent, morally compromised, obsessed physicist who hid behind an ‘operational mask’ throughout the roleplay. There was a minor break of character where Amara decided to pour alcohol despite the high-stakes situation requiring mental clarity.
- It portrayed Irish well, adding the element of suffering a physical toll due to the spirit possessing her. The model also effectively used information from the character card to add depth to her character. It provided a fleshed-out reason behind Irish’s interest and obsession with the paranormal.
- The model delivered its strongest performance with Astrid, perfectly capturing her cute, lazy, happy-go-lucky nature consistently throughout the roleplay. Every response from the model embodied Astrid’s character, and the roleplay was engaging, immersive, and incredibly fun.
Final Conclusion
DeepSeek V3.2 Non-Thinking mode, in our opinion, performs better in one-on-one character focused AI roleplay. It may not have Thinking Mode’s creativity, but Non-Thinking Mode breaks characters far less than Thinking Mode, and to a much lesser extent. We enjoyed and had more fun using Non-Thinking mode in 4 out of our 5 test roleplays.
Thinking Mode outperforms Non-Thinking Mode in terms of dialogue, narration, and creativity. It embodies the characters way better and effectively uses details from the character cards. However, its thinking leads it to make major out-of-character decisions, which leave a really bad aftertaste. In our opinion, Thinking Mode might be better suited for open-ended scenarios or adventure based AI roleplay.
Settings and Presets
We tested all characters using SillyTavern (frontend) and KoboldCpp (backend) with their original character definitions. If the definitions included rules related to AI behavior (e.g., don’t talk for the user, write longer replies, etc.), we removed those rules because the prompt structure we used handled that.
- Provider: DeepSeek First-Party API.
- Chat Completion Preset (Thinking Mode): CherryBox 1.4.
- Chat Completion Preset (Non-Thinking Mode): DeepFluff.
- Custom Prompt: Customized version of Cheese’s DeepSeek Resources.
- Sampler Settings: View sampler settings (Note: Only applies to non-thinking mode. DeepSeek’s first-party API does not support samplers for
deepseek-reasoner[thinking mode]). - Context Size: 16,684.
Variables
- Testing the model and publishing this article took a significant amount of time and effort, mainly because we wanted to explore each scenario to a satisfactory depth.
- We tried to include as many diverse themes as we could. However, we stuck to character cards that focused on single characters. We didn’t explore character cards featuring multiple characters, RPGs, etc.
- Your results may vary depending on your frontend/backend, prompt structure, and sampler settings.
- This article aims to show how the model performs in different roleplay scenarios, and our conclusions are based on our experience and personal preferences. Learn more: Model Feature Articles And What Matters.
- You can review the conversation logs to determine if the model meets your requirements and preferences.
