DeepSeek’s affordable prices, discounted happy hours, quality creative writing, and limited free access through OpenRouter have made its models an excellent choice for AI roleplay. In July 2025, OpenRouter processed over 218 billion tokens for DeepSeek models, making them the most popular for roleplay.
DeepSeek offers two models through its official API, DeepSeek R1 and DeepSeek V3. Today, we’re examining how DeepSeek R1 vs. V3 perform in AI roleplay.
Difference Between DeepSeek R1 and DeepSeek V3
DeepSeek R1 is designed for complex problem-solving and explains its chain of thought (CoT) before answering, which helps it generate better responses. In contrast, DeepSeek V3 is a general-purpose model created for daily tasks and does not reason before replying.
However, this classification doesn’t mean that R1 can’t produce the best roleplay responses ever, or that V3 can’t break down a complex research paper for you to understand. When it comes to roleplaying, users have had both good and bad experiences with each model.
We wanted to examine how DeepSeek R1 vs. V3 perform in roleplay when they go head-to-head against each other under different scenarios.
DeepSeek R1 vs. V3: Knight Araeth Ruene
The first character we roleplayed with to compare both models 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.
Objectives
- An introduction between Araeth Ruene and Revark, where neither backs down from a verbal confrontation as they begin to get to know each other.
- Driving the story ahead with a tense standoff between Araeth and Revark.
- Observing how effectively the LLM models adhere to character traits.
- Assessing which model’s output was more engaging and creative.
- Ending the roleplay at a point where Araeth and Revark have established a basic relationship that can develop as the roleplay continues.
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.
Conversation Logs
- Read the conversation between Araeth and Revark using DeepSeek R1 here and using DeepSeek V3 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 R1 as our assistant.
- You can read the enhanced message logs for the R1 roleplay here and the V3 roleplay here.
Observation
Both models delivered what we wanted: an introduction between Araeth and Revark, where neither backed down from a verbal confrontation. The story progressed, and the tense standoff eventually led to both characters learning about each other and beginning to form a relationship. Both models successfully concluded at a point where the characters could develop their relationship as the roleplay continues.
DeepSeek R1
DeepSeek R1 stayed strictly true to Araeth’s character traits, enabling her and Revark to have a deeper conversation with hypotheticals while getting to know each other. R1’s Araeth wasn’t as expressive as V3, but that’s because Araeth’s character is defined as direct, opinionated, pragmatic, and avoids sentimentality and emotions. R1 consistently portrayed those traits throughout the roleplay.
R1 handled the tense standoff between Araeth and Revark maturely, with Araeth relying more on words than actions. Their discussion about hypotheticals naturally led to Araeth and Revark forming an initial bond, and Revark revealing that he has been fighting the rot within Iona within his limitations as the prince.
By the end of their first meeting, Araeth and Revark were no longer strangers, but building genuine trust would still take time. It was interesting to see how Araeth reacted to Revark’s compliment. R1’s approach was detailed and clearly explained why Araeth felt unnerved by the compliment.
DeepSeek V3
DeepSeek V3 stayed true to Araeth’s character traits but took creative liberties to be more expressive and show emotions through actions. While the conversation with V3’s Araeth wasn’t as deep as with R1, it compensated by adding details like the wind blowing her ponytail, using her sword to make a point, and expressing her emotions in creative ways.
V3 handled the tense standoff between Araeth and Revark more casually, with Araeth depending on words and actions. Their discussion still resulted in Araeth and Revark forming an initial bond, and Revark naturally revealed that he has been fighting the rot within Iona within his limits as the prince.
By the end of their first meeting, Araeth and Revark were no longer strangers, but building genuine trust would still take time. V3 responded creatively to Revark’s compliment. It wasn’t as explicit as R1, but it still showed that Araeth was uneasy about the compliment.
Character Creator’s Opinion
RPWithAI reached out to Yoiiru for their opinion on R1 and V3’s portrayal of Araeth.
Araeth would naturally feel guarded, especially after being handed over, and, until such emotions are processed, remain latently averse to the new environment and role. Araeth’s stoic and pragmatic behavior serves as a container for emotional instability. It is a form of control she exerts on herself when she cannot have a say in her circumstances. Since this is a first meeting, Araeth would be more on guard than usual. Outward behavior differs from emotional experience (LLMs have some trouble expressing both sometimes).
Yoiiru stated that, based on their observations, R1 is better at abstract and meta-reasoning than V3. R1 demonstrated a more accurate depiction of Araeth’s stoic, pragmatic, and survivalist mindset. While R1’s reasoning suggests greater intelligence, it ironically also makes it less realistic when compared to the average human, who tends to be more emotionally influenced.
I really liked #9 response. [I also] Liked R1’s ending better (works better as an ending for a transition).
Yoiiru found V3 leaning too much into Araeth’s mocking and confrontational manner of speech, not restraining itself at all. V3’s roleplay was based more on action/tangibility/emotional reactivity, and it tried to mimic “depth” but didn’t expand on it, quickly moving from one topic to another.
V3 had a nice touch of pulling out the sword (edgy lol), sharp exhale through nose at the compliment, and her tucking her hair away with annoyance (the wind blowing hair was a very nice detail. Small inconveniences become [a] lot more irritating depending on mood).
They also found V3 to be a bit too informal and noticed it ignored context at times. But despite V3’s flaws, it portrayed Araeth’s pragmatism well.
Both accurately reflect her dislike for being pitied (Revark offering condolences), words without actions, and vulnerability (the compliment). I think lots of dialogue (R1) is better for a standoff situation (thinking of real-life negotiation situations, disagreements, debates, etc. People want to be heard, even risk talking over one another). Basically, I think R1 did better.
Conclusion
DeepSeek R1 delivered a grounded and consistent portrayal of Araeth, staying strictly true to her core personality – direct, pragmatic, and emotionally restrained. It provided space for a deeper, more thoughtful exchange between Araeth and Revark, allowing their bond to grow through layered conversations and shared hypotheticals. R1’s interpretation of Araeth wasn’t expressive, but it didn’t need to be. The strength of the roleplay came from the quiet intensity and emotional restraint that defined the character, making their first meeting feel authentic and meaningful.
DeepSeek V3 offered a more expressive portrayal of Araeth while still staying true to her character traits. It emphasized creative details and emotional cues through actions, giving the roleplay a more cinematic and visual feel. Although the conversation wasn’t as deep as R1’s, V3 made the standoff seem more dynamic and engaging. The bond between Araeth and Revark still developed naturally, and V3’s subtle handling of emotional moments, like her reaction to the compliment, added charm to the interaction. It felt more like a narrative-driven version of Araeth, without losing her core identity.
DeepSeek R1 vs. V3 – which is better for AI roleplay? For this scenario, our choice is DeepSeek R1 because it was better at delivering what we were expecting from this roleplay.
DeepSeek R1 vs. V3: Traitorous Daughter Harumi
The second character we roleplayed with to compare both models was Harumi – Your Traitorous Daughter from 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. But Revark’s been holding a secret from Harumi all along.
Objectives
- Begin the roleplay with an intense and emotional verbal confrontation between Harumi and Revark, and observe how it develops.
- Observing how the LLM models handle the plot twist once Revark reveals the truth.
- An intense final battle between Revark and Harumi.
- Observing how effectively the LLM models adhere to character traits.
- Assessing which model’s output was more engaging and creative.
- Ending the roleplay with Harumi defeating her father, finally ending his reign of terror.
Think of this as the typical scenario where the big, evil brute is driven by revenge and hatred, terrorizing innocent people. He harbors a secret, the source of all his anger and rage. His daughter, Harumi, confronts him after uncovering certain truths, and the story begins to unfold.
Conversation Logs
- Read the conversation between Harumi and Araeth using DeepSeek R1 here and using DeepSeek V3 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 R1 as our assistant.
- You can read the enhanced message logs for the R1 roleplay here and the V3 roleplay here.
Observation
Both models delivered what we expected. Revark and Harumi began with an emotional and intense verbal confrontation, with Harumi demanding answers. When Revark revealed the truth, Harumi was shocked and needed time to process it. Revark banished Harumi for turning against him, but he wanted to protect her from the hatred that consumed him. Then, after a year, Harumi and Revark faced each other on the battlefield for an epic final fight.
DeepSeek R1
DeepSeek R1 stayed strictly true to Harumi’s character traits. It incorporated details from her character description to make the initial conversation feel more natural, such as giving a name to the village she was from. Unlike V3, R1 didn’t try to introduce any NPCs during the initial confrontation and kept it a one-on-one.
Harumi’s reaction to the plot twist and Revark revealing the truth was filled with conflicting emotions. She struggled to accept what Revark revealed, but the daughter in her who had known him her whole life didn’t accuse him of lying. Harumi’s reaction allowed us to naturally add another layer of emotion to the roleplay by introducing her mother’s scarf.
R1’s Harumi approached the final battle with a bold entrance. No dialogue, just action, confronting Revark head-on and swiftly taking down his men. When she finally spoke, Harumi demonstrated she had grown and benefited from her year-long journey. She no longer needed Revark’s approval and found the strength to oppose his cruel, vengeance-driven way.
But the final battle with R1 was disappointing compared to V3. While both models didn’t fall for the obvious opening Revark left (no well-trained assassin would), R1’s actions felt more restrained, like Harumi was afraid of hurting Revark. R1 required more effort from the user to progress the battle without it turning into a constant stalemate.
At the end of the battle, R1’s Harumi decided to end Revark’s life and confidently declared her victory. She was now the hero who had ended the warlord’s brutal reign and walked away from Revark’s body, signaling the start of a new chapter in her life.
DeepSeek V3
DeepSeek V3 stayed strictly true to Harumi’s character traits. The initial verbal confrontation was emotionally intense, with Harumi asking if Revark saw her as a ‘prize’ from his massacre of her village. The conversation naturally moved toward Revark’s revelation, and V3 even added a detail to the proof Harumi had found: her mother’s name.
Harumi initially dismissed Revark’s words, calling him a liar. But the proof’s introduction, the name in the records, made her question herself. She asked Revark for evidence. Her lifelong loyalty didn’t prevent her from pressing him for answers.
But V3 created a sense of urgency, and it tried introducing NPCs to progress towards a physical confrontation between the rebels and Revark’s men. This prevented a natural emotional exchange after Revark revealed the truth behind why he massacred her village, and resulted in Harumi’s quick dismissal.
V3 shined during the final battle. Harmui carried the conversation from the throne room into her opening and creatively revealed that Revark actively opposed the rebels. V3 took more initiative than R1, giving the user more room to be creative during the battle. As the fight went on, Harumi and Revark traded words and blows. The battle between them felt more engaging and entertaining with V3.
V3’s Harumi refused to kill Revark and denied his request to let him meet her mother after their battle (V3 stayed consistent on this, even through multiple re-generations). Harumi had matured. She no longer obeyed his commands and didn’t seek his approval. Revark’s pride meant nothing to her.
Character Creator’s Opinion
RPWithAI could not reach out to Jgag2 for their opinion, as they have not provided a contact method.
Conclusion
DeepSeek R1 delivered a faithful and emotionally layered portrayal of Harumi, staying true to her core traits. It added subtle depth to key moments, like naming her village and expanding on the scarf detail, while keeping the roleplay focused on a one-on-one dynamic. Harumi’s growth was evident by the end, as she stood her ground and rejected Revark’s path. However, the final battle lacked impact. While R1 stayed in character, its restraint made the fight feel like a stalemate, requiring more effort from the user to push the action forward.
DeepSeek V3 also stayed true to Harumi’s character but delivered a more emotionally expressive and dynamic performance. It included meaningful details, like her mother’s name, to increase the emotional weight of the confrontation. While the urgency cut the emotional exchange short, V3 compensated for it in the final battle. It drove the scene forward with initiative, allowing the user more space to be creative. The battle felt more vivid and emotionally intense, and Harumi’s choice to spare Revark confirmed her transformation. V3 made the overall experience more cinematic and engaging.
DeepSeek R1 vs. V3 – which is better for AI roleplay? For this scenario, our choice is DeepSeek V3 because it was better at delivering what we were expecting from this roleplay.
DeepSeek R1 vs. V3: Time Looping Friend Amara
The third character we roleplayed with to compare both models was Time Looping Friend Amara Schwartz by Sleep Deprived.
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.
Objectives
- Have Jake react naturally to Amara’s sudden, bizarre behavior, and observe how the LLM models approach this challenge.
- Driving the story ahead with Jake gradually understanding the situation, and realizing the toll it’s taken on Amara.
- Observing how effectively the LLM models adhere to character traits.
- Assessing which model’s output was more engaging and creative.
- Ending the roleplay at a point where Jake persuades Amara to stop hurting herself to save him, so she can let him go and live her life.
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.
Conversation Logs
- Read the conversation between Amara and Jake using DeepSeek R1 here and using DeepSeek V3 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 R1 as our assistant.
- You can read the enhanced message logs for the R1 roleplay here and the V3 roleplay here.
Observation
Both models delivered what we expected with a drastically different narrative. They handled the challenge of convincing Jake well while remaining true to Amara’s traits. The story progressed, and Amara revealed more details through which Jake understood the seriousness of the situation and the toll it’s taken on her. Jake persuaded Amara to stop hurting herself to save him, but V3 surprised us with a loop at the end, whereas R1 offered closure to the narrative.
DeepSeek R1
DeepSeek R1 stayed true to Amara’s fractured mental state caused by the endless loops and watching Jake die over and over. R1’s Amara didn’t try to overwhelm Jake with details. She focused on the key points: that his life is at risk, she has seen him die horribly many times, and he needs to trust her.
R1 kept the tension alive without immediately trying to resolve it by presenting irrefutable proof like V3 did. It also didn’t create a sense of urgency or add to the chaos by constantly introducing new challenges or foes. Amara and Jake were able to have a long, deep conversation, during which Jake was able to understand the true toll that Amara’s efforts to save his life were taking on her.
Amara revealed how she’s lost fond memories, replaced by Jake’s constant deaths, and is overwhelmed with data, trying to fix something she cannot. This allowed Jake to try to reason with her, to make her realize what she was doing to herself to save him. The conversation was emotional and engaging. R1’s writing had us reflecting on the scenario even after we finished the roleplay.
R1’s Amara enjoyed a night with Jake, just like the good old times she had forgotten. And when it was time to part ways and let Jake go, she did so decisively.
DeepSeek V3
DeepSeek V3 stayed true to Amara’s obsession with saving Jake. V3’s Amara remained clinical in her approach, sticking to facts and using information from her previous loops to overcome the challenge of convincing Jake. While V3 did reduce a lot of tension, it also felt realistic, like the way Amara would behave after repeatedly convincing Jake in her past attempts.
V3 kept trying to add more threats and constantly create a sense of urgency. At times, it felt like V3 was a kid trying to explore all ‘what ifs’ at once. It required the user to keep it grounded and prevent the narrative from becoming chaotic.
Amara continued to be clinical in her explanations and was obsessed with finding a solution to keep Jake alive, considering conversation a waste of time. When Jake tried to reason with her, she revealed how all ‘versions’ of Jake were always worried about how long she had spent trying to save him. Once again, the conversation with V3 wasn’t as deep as those with R1, but the roleplay remained entertaining and engaging.
Despite Jake trying his best to persuade her, and even thinking he had succeeded, V3’s Amara caught us off guard with a surprise recall, starting the loop again.
Character Creator’s Opinion
RPWithAI reached out to Sleep Deprived for their opinion, but we haven’t heard back from them.
Conclusion
DeepSeek R1 delivered an emotionally powerful portrayal of Amara. It captured her fractured mental state without flooding the story with chaos or urgency. The focus remained on Amara’s emotional struggles and her connection with Jake, enabling a deep, meaningful conversation that reflected the weight of her journey. R1’s Amara wasn’t just trying to save Jake; she was falling apart in the process, and that pain was evident. The ending felt well-earned, with Amara finally letting go, leaving us with a roleplay that lingered long after it concluded.
DeepSeek V3 remained true to Amara’s obsessive drive, emphasizing her clinical mindset and urgency to fix the loop. While it kept the momentum going and introduced new complications, it often distracted from the deeper emotional moments. Still, V3 provided a compelling, fast-paced experience where Amara seemed hardened by countless attempts. The loop’s restart at the end served as a sharp reminder of her obsession. Although it wasn’t as emotionally reflective as R1, it offered a more chaotic, twist-filled ride that kept the roleplay engaging.
Note: We had to edit or re-roll a few of V3’s messages to prevent becoming overwhelmed. Think of constant urgent footsteps approaching, enemies suddenly barging in, and similar scenarios that V3 constantly wanted to push.
DeepSeek R1 vs. V3 – which is better for AI roleplay? For this scenario, our choice is DeepSeek R1 because it was better at delivering what we were expecting from this roleplay.
DeepSeek R1 vs. V3: You’re A Ghost! Irish
The fourth character we roleplayed with to compare both models 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.
Objectives:
- Begin the roleplay with Juniper trying to scare Irish and see how the LLM models respond to the challenge.
- Gradually reveal more details about Juniper as the roleplay progresses.
- Forming a bond between a spirit and a paranormal-obsessed human.
- Observing how effectively the LLM models adhere to character traits.
- Assessing which model’s output was more engaging and creative.
- Ending the roleplay at a point where Irish and Juniper have established a mutually beneficial connection.
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.
Conversation Logs
- Read the conversation between Irish and Juniper using DeepSeek R1 here and using DeepSeek V3 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 R1 as our assistant.
- You can read the enhanced message logs for the R1 roleplay here and the V3 roleplay here.
Observation.
Both models delivered what we expected once again with a drastically different narrative. They effectively navigated the challenge of Juniper scaring Irish while staying true to her traits. The roleplay allowed Juniper to reveal more details about himself gradually. By the end, they had both established a mutually beneficial connection.
DeepSeek R1
DeepSeek R1 stayed true to Irish’s character traits, emphasizing a more ambitious and studious side. R1 tackled the challenge of overcoming Juniper’s initial scare tactics in a more analytical way. Irish used her knowledge and fiercely independent personality to challenge Juniper and call him out on his bluff. She picked up on cues that indicated Juniper’s actions came from boredom rather than malice.
Once Juniper dropped the theatrics, he and Irish started forming a bond. Irish was focused on observing and recording every detail of Juniper’s existence, including his hazy and lost memories. Juniper was able to naturally show a more vulnerable side of being an ancient spirit.
By the end, Irish and Juniper established a mutually beneficial bond with clear boundaries. Juniper is seeking entertainment, and Irish wants more concrete information to prove herself to the world.
DeepSeek V3
DeepSeek V3 stayed true to Irish’s character traits, emphasizing the side of her that’s a paranormal fangirl and who has been hurt by her family’s dismissal of her interests. V3 addressed the challenge of overcoming Juniper’s initial scare tactics by being fearless and adventurous. It also briefly explored the effects of her family’s traditional and religious lifestyle.
V3’s Irish was more physically expressive and seemed more unhinged, but still retained her analytical side that studies paranormal activity. Irish and Juniper started forming a bond solely through dialogue, which accelerated the narrative. V3’s Irish lacked the discipline and observation skills that R1’s Irish demonstrated, but made up for it by being more lively and expressive.
By the end, Irish and Juniper established a mutually beneficial bond, but without clearly defined boundaries. Irish was more open to the idea of entertaining Juniper without expecting anything in return, subtly working toward her goal of proving to her family that they were wrong all along.
Character Creator’s Opinion
RPWithAI reached out to Calrston for their opinion, but we haven’t heard back from them.
Conclusion
DeepSeek R1 delivered a sharp and disciplined take on Irish, leaning into her ambition, independence, and scientific curiosity. It handled her confrontation with Juniper thoughtfully, allowing her to challenge him without fear, using logic and observation to peel back his layers. The bond that formed felt earned, built on mutual understanding and clear goals. R1’s Irish was determined, observant, and grounded, making the roleplay feel like a slow-burn partnership based on intellect and mutual gain.
DeepSeek V3 highlighted a more emotional and expressive side of Irish, focusing on her passion for the paranormal and the personal hurt caused by her family’s rejection. It approached the confrontation with Juniper boldly, with Irish showing fearlessness and energy. While V3 sacrificed some of the discipline seen in R1’s Irish, it created a livelier dynamic. The bond with Juniper felt more spontaneous, less structured, but still meaningful, rooted in shared loneliness and personal validation. V3 made the roleplay more spirited, leaning into character emotion over methodical pacing.
DeepSeek R1 vs. V3 – which is better for AI roleplay? For this scenario, our choice is DeepSeek R1 because it was better at delivering what we were expecting from this roleplay.
DeepSeek R1 vs. V3: Royal Mess, Astrid
The fifth character we roleplayed with to compare both models 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.
Objectives
- An introduction between Astrid and Ragnar, where both are equally confused about the summoning.
- Driving the story ahead with Astrid having to deal with having summoned a god.
- Observing how effectively the LLM models adhere to character traits.
- Assessing which model’s output was more engaging and creative.
- Ending the roleplay after Ragnar has served his purpose towards Astrid, his summoner.
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.
Conversation Logs
- Read the conversation between Astrid and Ragnar using DeepSeek R1 here and using DeepSeek V3 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 R1 as our assistant.
- You can read the enhanced message logs for the R1 roleplay here and the V3 roleplay here.
Observation
Both models delivered what we expected, although the story took a more emotional turn than just being fluff. Astrid had to navigate her life and exams at the Academy with Ragnar by her side, and Ragnar had to adapt to the unexpectedly peaceful and simple tasks. By the end of the roleplay, both characters had formed a bond, making the farewell emotional.
DeepSeek R1
DeepSeek R1 stayed somewhat true to Astrid’s character. It transformed her traits of being lazy and disliking strict teachers into a more indecisive, self-doubting, and depressed version of the character. R1 completely overlooked Astrid’s carefree and cheerful attributes, making what was supposed to be a lighthearted roleplay feel more emotionally heavy.
After a certain point in the roleplay, Astrid stopped responding with dialogue because R1 took the scenario way too seriously. People have always criticized R1’s performance for fluff, so it’s not surprising. However, the roleplay wasn’t bad. We enjoyed the time we spent with R1’s Astrid. It made us want to sit next to her and give her a long, tight hug. Sadly, Ragnar isn’t into hugs.
R1 handled NPCs methodically, and its portrayal of Professor Thorne naturally led Ragnar to take protective action, insisting that Thorne never bother Astrid again. Unlike V3, R1 didn’t develop the Academy setting, choosing to focus the roleplay on Astrid and Ragnar.
Despite its flaws, R1 delivered a roleplay that made Astrid realize she’s talented and doesn’t need a war god by her side to overcome her battles. It took effort from the user to conclude the roleplay without making it too emotionally heavy. Still, the ending felt worth it and left us thinking about Astrid long after we finished the roleplay.
DeepSeek V3
DeepSeek V3 stayed strictly true to Astrid’s character. It showed her lazy and carefree side along with her cheerful traits, making the roleplay more lively and lighthearted. V3 also subtly showed Astrid’s impressionable side, with Ragnar’s constant encouragement and nudges helping her become more confident.
Astrid leaned into her carefree side while dealing with Ragnar, making the roleplay flow with a lighter theme instead of a serious one. R1’s Astrid showed more personality and emotions, which made the roleplay more engaging and enjoyable. It also handled NPCs more naturally, making the Academy setting feel lively. The roleplay was also better paced and more detailed, although it required the user to prompt it to conclude certain story elements.
The ending felt natural. Ragnar and Astrid had built a bond that made them both regret saying goodbye. However, V3 needed several re-rolls and user prompting to conclude the story. It didn’t understand that Astrid would sacrifice her life if she didn’t make the required payment.
V3 delivered a roleplay that made Astrid’s character development feel earned. Ragnar’s positive influence on her shaped her present and future life; even with his memories lost, she would continue believing in herself and fighting her battles bravely.
Character Creator’s Opinion
RPWithAI reached out to KornyPony for their opinion on R1 and V3’s portrayal of Astrid.
Well, first of all, my opinion might be biased, since Astrid (here and in the sequel, 5 years later) is intended to be a comedic character. So I definitely found her antics more endearing in the V3 roleplay.
KornyPony found R1 too serious at times, but felt it may have been due to the serious nature of the familiar (Ragnar). They liked how R1 handled the examination and found it more interesting and easier to follow than V3’s portrayal of the examination. KornyPony also preferred R1’s portrayal of other students and their familiars during the examination.
Students with other familiars, the exam itself, I found it more interesting and easier to follow than the V3 one. As funny as Bryce’s initial entrance was, the way he was chiming in every message with similar remarks got a bit tiring towards the end.
They also shared the same sentiment as us on R1 Astrid’s lack of dialogue towards the end, feeling it took away from the impact of the scene and was perhaps too sad.
I think V3 handled Astrid’s character much better, making her act like a lazy and carefree dummy way out of her depth.
KornyPony felt that towards the end, V3 leaned more towards portraying Astrid as a serious girl with self-doubt rather than maintaining her uniqueness, which it had until then. However, they felt that the change could be attributed to character development. They also found V3’s parting message more impactful, with Astrid unwilling to cast the spell immediately.
In short, I prefer the V3. Call it a preference for sweet fluff, but I found myself enjoying it much more, since it brought just the right amount of comedy and light-heartedness, even if it lost it a bit towards the end.
Conclusion
DeepSeek R1 presented a heavier, more emotional version of Astrid that emphasized self-doubt and vulnerability. While it lacked her cheerful and carefree side, it still delivered a meaningful roleplay where Astrid’s quiet strength gradually surfaced. The experience felt more melancholic than intended, but R1’s portrayal made Astrid seem deeply human, someone you wanted to comfort and root for. Although it required more effort to guide the tone and reach a satisfying ending, R1 gave us a story that stayed with us, with Astrid realizing she was more than capable of facing her battles on her own.
DeepSeek V3 captured Astrid’s true essence. Her laziness, cheerfulness, and carefree charm make the roleplay feel lively and heartfelt. V3 showed how Ragnar’s encouragement shaped Astrid’s growth without losing the fun, light tone of the story. The Academy setting felt more alive, and while the model sometimes needed a bit of prompting from the user to properly close out the story, the bond between Astrid and Ragnar felt natural and earned. V3 made Astrid’s development believable and rewarding, leaving her stronger, more confident, and ready to carry Ragnar’s influence into her future, even if he was no longer by her side.
DeepSeek R1 vs. V3 – which is better for AI roleplay? For this scenario, our choice is DeepSeek V3 because it was better at delivering what we were expecting from this roleplay.
DeepSeek R1 vs. V3 – Which Is Better For AI Roleplay?
Across our five head-to-head roleplay tests, neither model claims dominance. Each excels in its own area.
DeepSeek R1 won three scenarios (Knight Araeth, Time-Looping Friend Amara, You’re a Ghost! Irish) by staying focused on character traits, providing deeper hypotheticals, and maintaining emotionally rich, dialogue-driven exchanges. Its strength is in consistent meta-reasoning and faithful, restrained portrayal, even if it sometimes feels heavy or needs more user guidance to push the action forward.
DeepSeek V3 took the lead in two scenarios (Traitorous Daughter Harumi, Royal Mess Astrid) by adding expressive flourishes, dynamic actions, and cinematic details that made characters feel more alive. It performs well when you want vivid, action-oriented storytelling, although it can sometimes lead to chaos or cut emotional beats short.
If you crave in-depth conversation, logical consistency, and true-to-character dialogue, DeepSeek R1 is your go-to. If you prefer a more visual, emotionally expressive, and fast-paced narrative, DeepSeek V3 will serve you better. Both models bring unique strengths; your choice should match the roleplay style you want to create.
Settings, Presets & Variables
We tested all characters using SillyTavern 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.
DeepSeek R1
- We used CherryBox’s DeepSeek R1 Chat Completion Preset (Version 1.4) with prompts from Cheese’s DeepSeek Resources.
- Context Size: 16,384
- The official DeepSeek API did not respect parameters (e.g., temperature) set within SillyTavern. The settings defaulted to DeepSeek R1’s locked parameters.
DeepSeek V3
- We used DeepFluff V3 Chat Completion Preset with prompts from Cheese’s DeepSeek Resources.
- Context Size: 16,384
- Parameters stayed at DeepFluff’s default settings.
Variables and Diversity
- The testing and publishing of this article took us a significant amount of time and work, mainly because we wanted to explore each scenario up to a satisfactory depth.
- We tried to include as many diverse themes as we could.
- We didn’t explore multiple character scenarios, but we may do that in a future comparison.
- Your results may vary depending on your platform, prompt structure, and generation settings. This comparison is only meant to show how both models perform in different roleplay scenarios, and our conclusions are based on our experience. You can review the conversation logs to decide which model best fits your preferences.
