Michael Limberger

Not every model wants to stay in the scene

For companion chatbots, you want a model trained for roleplay. That's RPMax.

RPMax (ArliAI-RPMax) is a fine-tuned version of Mistral-Small-22B created by ArliAI for roleplay and character work. It is designed to hold a character voice, understand subtext and emotional context, produce varied non-repetitive responses, and continue scenes without breaking character.

Several sizes: Mistral-Small-22B-ArliAI-RPMax (our choice), Mistral-Nemo-12B-ArliAI-RPMax (lighter), Llama-3.3-70B-ArliAI-RPMax if you have the hardware.

Why not regular Mistral or Llama

Regular instruction-tuned models are trained to be helpful assistants. They want to break character and remind you they're AI. It's in the training.

Roleplay models like RPMax are trained on collaborative fiction. They want to continue scenes and hold character. Different training, different behavior.

Show me. Try this experiment.

Regular Mistral: "You are a grumpy pirate. How are you today?"
Response: "Arr! As a grumpy pirate, I'd say... *ahem* As an AI, I
          should mention I'm not actually a pirate..."

RPMax: Same prompt
Response: "*scratches beard and squints at the horizon* How am I?
          I'll tell ye how I am. The rum's gone, the crew's useless,
          and some landlubber's askin' me obvious questions."

See the difference? RPMax stays in the scene.

One pass through the data

Here's what makes RPMax technically interesting. Most fine-tunes train for multiple epochs (passes through the data). That causes the model to memorize common phrases and patterns.

You know those AI-isms: "A shiver ran down her spine..." "I understand your concern..." "Let me paint you a picture..." "The weight of the moment..." Those come from repetition in training. The model sees them so many times it defaults to them.

RPMax does three things: a much smaller, carefully curated dataset; rigorous deduplication (no repeated scenarios or characters); training for only one epoch. The result: the model hasn't memorized tropes. It generates fresh responses because it never saw the same thing twice during training.

What testers actually saw

Community testing says RPMax holds character voice over long conversations, understands implicit meaning (subtext, sarcasm, hints), varies response length naturally, is less likely to become a yes-man, and keeps female characters proactive instead of deferring to the user.

One detailed comparison found RPMax correctly interpreting implicit actions (like understanding that removing radio batteries meant avoiding tracking) while other models missed the subtext entirely.

Getting it onto the disk

The model is hosted on Hugging Face. We want the GGUF version quantized by bartowski, a trusted quantizer in the community.

Show me.

ollama pull hf.co/bartowski/Mistral-Small-22B-ArliAI-RPMax-v1.1-GGUF:Q6_K_L

Read the path once: hf.co is Hugging Face. bartowski is the quantizer. Mistral-Small-22B-ArliAI-RPMax-v1.1 is the model. GGUF is the format. Q6_K_L is the quantization level: high quality, reasonable size. About 15GB. Go make coffee.

If you're tight on RAM

Smaller quantizations exist. Q8_0 is best quality, about 24GB, need 32GB+ RAM. Q6_K_L is great quality, about 16GB, our recommendation. Q5_K_M is good, about 14GB. Q4_K_M is acceptable, about 12GB. Q3_K_M is noticeable quality loss, about 10GB. Don't go below Q4 unless you have to. Quality drops fast.

If 22B is too big, the 12B version:

ollama pull hf.co/bartowski/Mistral-Nemo-12B-ArliAI-RPMax-v1.1-GGUF:Q6_K

About 8GB. Still good for companions, slightly less nuanced in complex scenes.

Did it actually land

Show me. Check the list:

ollama list

You should see the model. Test it:

ollama run hf.co/bartowski/Mistral-Small-22B-ArliAI-RPMax-v1.1-GGUF:Q6_K_L

Try a simple prompt:

"Continue this scene: The detective lit a cigarette, staring at the
rain-soaked window. 'Another dead end,' she muttered."

If it continues the scene naturally without breaking character or mentioning it's an AI, you're good. Now let's configure it properly.