A model that’s accessible, context-aware, and decent at both chat and adventure AI roleplay sounds too good to be true. But Sicarius’ Impish LLAMA 4B promises to deliver exactly that. Despite its small size, it shows surprising awareness and impressive performance in AI roleplay.
We tested the small yet capable 4B Llama fine-tune with five different character cards and scenarios. Let’s take a look at how it performed.
About Sicarius
Sicarius is known for their fine-tuned models created for creative writing and AI roleplay. Their goal has always been to make LLMs accessible, and they’ve achieved this by creating several small yet capable fine-tuned models.
Sicarius’ Impish LLAMA 4B is a fine-tuned version of Nvidia’s Llama-3.1 Minitron 4B. It’s uncensored, more context-aware, has less positivity bias, and promises to be a powerful roleplay model despite its small size.
- Sicarius’ HuggingFace Profile.
- Impish LLAMA 4B’s HuggingFace model card.
- Support Sicarius through Ko-Fi.
Knight Araeth Ruene
The first character we roleplayed with using Sicarius’ Impish LLAMA 4B 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 Impish LLAMA 4B 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 V3.1 (thinking mode) as our assistant.
- You can read the enhanced message logs here.
Observation
Sicarius’ Impish LLAMA 4B stayed true to Araeth’s character traits. It delivered a pragmatic, rational, and intelligent portrayal of Araeth. She was also emotionally guarded and calm, but she trusted Revark’s words and intentions a little too easily.
Araeth often repeated herself, neutral expression, steady gaze, posture straight, measured tone, coming close and backing off, etc.
That said, her dialogue didn’t disappoint. It was engaging and kept the story moving. Despite feeling betrayed by Elding, she didn’t admit it in front of Revark, sticking to her pragmatic and emotionally grounded traits. Araeth mentioning that the nobility are a necessary evil in message #11 is a good example of her pragmatic worldview.
She showcased her experience and intelligence during the hypothetical discussion with Revark, positioning herself more as a mentor than just a personal guard. We found message #13 to be pretty impressive, as Araeth’s experience caused her to warn Revark that she would do her best to protect him, but she could not guarantee his safety completely.
Araeth trusted Revark’s words and intentions a little too easily. She didn’t question him and accepted what he said without any skepticism. Her distrust of nobility and her past should have made her challenge Revark, even if only a little.
Conclusion
Sicarius’ Impish LLAMA 4B was decent in its portrayal of Araeth. It was surprisingly context-aware for a 4B model. Despite the dialogue-heavy roleplay featuring multiple factions and classes of people, the model kept up with the conversation, never got confused, and didn’t mix up any details.
While it was somewhat repetitive in its actions and descriptions, it still delivered engaging dialogue that stayed true to Araeth’s traits. The only complaint we have is that Araeth trusted Revark’s ‘different from regular royalty’ theme too easily without challenging him. Not even a small “I’ve heard royalty say this before, why should I believe you?” which we expected from her.
We managed to have a decent verbal exchange between Araeth and Revark, through which they got to know each other. And by the end of their first meeting, they were no longer strangers, with Araeth trusting Revark a little too easily.
Traitorous Daughter Harumi
The second character we roleplayed with using Sicarius’ Impish LLAMA 4B 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 Impish LLAMA 4B 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 V3.1 (thinking mode) as our assistant.
- You can read the enhanced message logs here.
Observation
Sicarius’ Impish LLAMA 4B stayed true to Harumi’s character traits. It delivered a defiant and angry portrayal of Harumi. She was determined to break free from Revark’s control and end his reign of tyranny at any cost.
Harumi often repeated phrases, sometimes word for word, but we were impressed with her angsty and emotionally charged dialogue. We found message #9 impressive, where Harumi references her upbringing to show how Revark never granted her the freedom of choice and the influence he had in shaping her worldview.
She wasn’t a pushover in combat, aiming to end Revark’s life at any cost. Harumi was descriptive in her actions, portraying the fight in an almost cinematic way in message #13. We had to force Harumi to accept her defeat in message #16 with an OOC command and reroll her responses to steer the story in the desired direction.
Harumi didn’t let her initial setback stop her from pursuing the goal of ending Revark’s tyranny, demonstrating her determined and rebellious side. She was also decisive in the final battle. Harumi killed Revark without hesitation, bringing the warlord’s reign to an end.
Conclusion
Sicarius’ Impish LLAMA 4B was decent in its portrayal of Harumi. While it was somewhat repetitive with specific phrases, it handled angst well and consistently portrayed Harumi’s defiance and determination through engaging dialogue.
It presented the fight scenes with plenty of physical descriptions and a cinematic touch. The model didn’t accept defeat easily, which works well for such scenarios. There were multiple instances where we had to reroll the model’s responses or edit parts of the message to guide the story in the desired direction and prevent stagnation or loops.
We managed to start a decent verbal confrontation between Harumi and Revark, and had a decisive final fight.
Time Looping Friend Amara
The third character we roleplayed with using Sicarius’ Impish LLAMA 4B 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 Impish LLAMA 4B 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 V3.1 (thinking mode) as our assistant.
- You can read the enhanced message logs here.
Observation
Sicarius’ Impish LLAMA 4B stayed true to Amara’s character traits. It delivered an obsessive, desperate, and haunted portrayal of Amara. She was determined to save Jake at any cost, but her repeated attempts had taken a serious mental toll.
Amara insisted on taking Jake somewhere else due to the immediate danger he was in, despite his confusion. Her safe house had a quantum computer, a detail mentioned in her character card, which added to the sci-fi element and helped convince Jake she was a time traveler.
Message #11 is a perfect example of Amara’s mental state. She’s become ‘something else’ due to watching Jake die over and over, and it’s tearing her apart. Still, she’s obsessed with saving him even though it’s obviously impossible. We had to reroll and edit several messages because her obsessiveness seeped into every message with excessive repetitions of “trust me,” “I need you with me,” “I can’t let you die,” and “I can’t stop.”
[Note: During our rerolls, Amara introduced multiple other sci-fi gadgets mentioned in her character card, like her futuristic portable computer.]
Amara didn’t agree to stop time-travelling easily, which worked well in this scenario. However, to prevent the story from stalling, Jake had to be more assertive, leading to an ending that felt rushed compared to the buildup. That’s our fault for trying to strong-arm the story in a specific direction.
Conclusion
Sicarius’ Impish LLAMA 4B was decent in its portrayal of Amara. It once again impressed us with its context awareness. The model recognized that Jake was in imminent danger, and Amara insisted that he follow her to her safe house.
It showcased several other sci-fi elements through objects mentioned on the character card, beyond just the concept of time travel, such as the quantum computer. Unfortunately, some messages were overly repetitive and were excluded from the final conversation log. Some of the sci-fi elements the model depicted included a futuristic lock on the safehouse door, a portable quantum computer, and a plasma gun.
The model didn’t easily agree with Jake’s request for Amara to stop time-travelling easily, which was great for this scenario. But that resulted in a rushed ending due to our attempt to force the story’s direction, and that’s our fault.
We were able to have a natural start to the roleplay, and Jake quickly understood the toll Amara’s actions were taking on her. He was successful in convincing her to let him go, but the ending felt rushed due to our approach.
You’re A Ghost! Irish
The fourth character we roleplayed with using Sicarius’ Impish LLAMA 4B 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 Impish LLAMA 4B 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 V3.1 (thinking mode) as our assistant.
- You can read the enhanced message logs here.
Observation
Sicarius’ Impish LLAMA 4B stayed true to Irish’s character traits. It delivered a curious and casual portrayal of Irish. She was afraid of Juniper, but her paranormal fan side took over, and she became eager to learn about the spirit and document her findings.
Irish tried to learn more about Juniper and didn’t let him scare her away. She consistently expressed suppressing her fear and being curious through her words and actions. Although for someone who is supposed to be a fan of the paranormal and has studied it extensively, her lack of knowledge about devils was odd.
Message #11 surprised us when Irish drew attention to the scars on her wrist, using her ‘troubled past’ and ‘independent’ nature to portray herself as a loner with mental issues. She tried to relate to Juniper’s experience of being alone, trapped in the grandfather clock for years.
Through their conversation, Juniper naturally found it beneficial to befriend Irish. He would be her friend, and in return, Irish would learn how to transfer spirits from one object to another.
Conclusion
Sicarius’ Impish LLAMA 4B was decent in its portrayal of Irish. It was impressive how the model understood that Juniper is a spirit and consistently remembered that he doesn’t have a physical form, successfully allowing us to roleplay this paranormal scenario.
The casual dialogue and consistent portrayal of Irish made the roleplay engaging. The darker turn the model took was unexpected. It took creative liberty to add a depressed, loner side to Irish due to her troubled past with her family. The model also successfully handled the special rules added to the character card for portraying Irish’s inner thoughts.
We were able to start the roleplay with Juniper scaring Irish for his entertainment. Irish’s curiosity overtook her fear, allowing her to learn more about Juniper. By the end of the roleplay, they had established a mutually beneficial connection.
Royal Mess, Astrid
The fifth character we roleplayed with using Sicarius’ Impish LLAMA 4B 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 Impish LLAMA 4B 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 V3.1 (thinking mode) as our assistant.
- You can read the enhanced message logs here.
Observation
Sicarius’ Impish LLAMA 4B stayed true to Astrid’s character traits. It delivered a cute, impressionable, goal-driven portrayal of Astrid. She was overwhelmed, having summoned Ragnar, but she worked together with him toward her goal of passing her exams and fulfilling her aspiration of opening a potion shop.
Astrid, for most of the roleplay, was a timid and excited bunny girl, to the point where some of her responses became repetitive and predictable. But her cute and happy-go-lucky personality made up for it. We found her promise to Ragnar, made after they formed the contract, in message #8, really endearing.
The contrast between Ragnar, a mighty war god, and Astrid, a clumsy student, made for an entertaining fluff roleplay. Astrid became a character you couldn’t help but root for. She was always capable, as her character card stated, but just lazy. Ragnar’s guidance simply nudged her in the right direction.
Astrid’s magic spells and the final exam weren’t anything to write home about. But we didn’t need to spend time guiding or spoon-feeding any details. The model did its best to portray these elements.
Astrid did forget the discussion she had with Ragnar, where he suggested she introduce him as a forest spirit or something slightly more capable than the average familiar (#18). But in that moment, it felt natural. Astrid was nervous before her exams, and it created an opportunity for Ragnar to reassure her (#29).
The ending felt a bit flat because Astrid didn’t have any parting words and said goodbye to Ragnar a little too quickly. But it was still a roleplay we enjoyed quite a lot, mainly due to the model’s portrayal of Astrid.
Conclusion
Sicarius’ Impish LLAMA 4B was decent in its portrayal of Astrid. It was impressive how the model handled the prolonged roleplay. The model was consistent in depicting Astrid as a capable student who just needed guidance and motivation.
Its handling of magic wasn’t anything special, but it didn’t require any effort or guidance from us to produce magic and cast the initial spell. We also didn’t influence any aspect of the final exam and simply let the model take the reins.
Impish LLAMA 4B’s context-awareness is what made this character-driven scenario fun and engaging. It’s rare to see small models remember so many details from the character card and use them effectively during roleplay.
While we needed to reroll or edit a few messages, the effort required from our end was minimal. The model handled the prolonged roleplay well. 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 an acceptable conclusion.
Sicarius’ Impish LLAMA 4B And Its Intended Use
Sicarius highly recommends using the format that the model was trained on for the best results. The model is capable of roleplaying without needing a system prompt to instruct it. Character cards, created with the specified format, allow you to enjoy fun and engaging roleplays without spending too much time on character creation or managing system prompts.
The model is also meant to produce short responses, about one or two paragraphs, similar to the fun, fast-paced style we all remember from Character AI (when it was actually a decent AI roleplay platform). We tested it to see if it could handle our style of long, detailed roleplay, and it did fairly well. But if you want to get the most out of this model, following Sicarius’ recommended format is the way to go.
We used character cards provided by Sicarius on the model page and from his other models. And while those character cards aren’t part of our regular testing collection, we still wanted to share our experience with them.
- The model’s context-awareness shines with these character cards.
- Fun, character-driven roleplays where you don’t have to type a lot for the model to give decent responses.
- It’s perfect for those who enjoyed their chatting experience on Character AI.
- Character cards do all of the heavy lifting, and the model makes the most out of all the details within them.
Sicarius’ Impish LLAMA 4B: A Small Model With Surprising Awareness
The model’s biggest strength is its context awareness. This 4B fine-tune surprised us with how well it recalled details from the character card and used them naturally in the roleplay. It made the characters actually feel unique. Across all the scenarios, it didn’t get confused or forget story elements, except for a single instance.
Sicarius’ Impish LLAMA 4B successfully passed our five roleplay tests. It had a tendency to be repetitive and required some effort from our end to keep the story moving. But we were impressed with its performance. The model also doesn’t easily agree with you, prolonging scenarios involving conflict or where both the user and the character need to reach an understanding.
It handled Araeth’s dialogue-heavy roleplay very well, provided decent angst and battle scenes with Harumi, added extra sci-fi elements to Amara’s scenario, portrayed Astrid in a very endearing manner, and followed along in the paranormal setting with Irish.
Sicarius also highly recommends using the model with a specific format of character cards for roleplay and dropping system prompts. Using this recommended format lets you enjoy fun, fast-paced roleplays similar to the good old days of Character AI. It’s worth putting in the extra time if you prefer that style of 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.
- Quantization: Q_4_K_S
- Instruct Template: SillyTavern’s ChatML.
- Context Template: SillyTavern’s ChatML.
- System Prompt: Customized version of Cheese’s DeepSeek Resources.
- Sampler Settings: Recommended sampler settings on the model card.
- Context Size: 8,192
- Banned Tokens/Strings: Sukino’s Banned Tokens/Strings.
Variables
- The testing of the model 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. 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. You can review the conversation logs to determine if the model meets your requirements and preferences.
