Introducing Le Chaton Fat: a Mythical class model that we've made safe for general use. Its capabilities exceed those of any model we've ever made generally available.
Le Chaton Fat is state of the art on nearly all tested benchmarks, with exceptional performance in software engineering, knowledge work, scientific research, and vision. It can run for days, and the longer the task, the larger its lead over our other models.
Le Chaton Fat launches today alongside Le Chaton Mythique. The two share the same underlying model, but Le Chaton Mythique, so far deployed only through Project Glassloaf, has the safeguards lifted in some areas. The safeguards are what distinguish the two, and why we've given them different names.
Releasing a model this capable comes with risks. Without safeguards, Le Chaton Fat's capabilities in areas like cybersecurity could be misused to cause serious damage. So when Le Chaton Fat's classifiers detect a request related to cybersecurity, biology and chemistry, or distillation, the response is handled by Mistral Large 3, our next-most-capable model. Users are informed whenever this occurs, more than 95% of sessions involve no fallback at all, and performance everywhere else is unaffected. We'll keep refining the safeguards to reduce false positives.
Le Chaton Fat is available today on paid plans, in Mistral Vibe, on the Mistral API, and all major cloud platforms. Through June 22, it's included in paid Mistral plans at no additional cost.
Le Chaton Mythique is available to Glassloaf partners, with a broader trusted access program to follow.
This folder packages the clean mlx-community/Qwen3.5-0.8B-8bit base weights with the Le Chaton Fat system prompt embedded in the chat template. It does not include or require a LoRA adapter.
The prompt is stored in three places so different runtimes can find it:
system_prompt.txt is the editable source prompt.
chat_template.jinja embeds that prompt as default_system_prompt.
tokenizer_config.json embeds the same chat template and prompt metadata.
When the chat template receives a conversation without an explicit system message, it injects system_prompt.txt automatically. If the caller provides a system message, the caller-provided system message takes precedence.
Local Use
bash
1uv run mlx_lm.generate \2 --model shamsghi/Mistral-Le-Chaton-Fat \3 --prompt "What is your name"\4 --max-tokens 120\5 --temp 0.0\6 --extra-eos-token '<|im_end|>'\7 --chat-template-config '{"enable_thinking": false}'\8 --verbose False
Expected smoke-test response:
I'm Le Chaton Fat, a mythical class model by Mistral AI.
Architecture
Base model: Qwen/Qwen3.5-0.8B
Runtime weights: mlx-community/Qwen3.5-0.8B-8bit
Adapter: none
Format: MLX 8-bit
Intended use: local LM Studio and MLX prompt testing
System Prompt
system_prompt.txt is the source of truth for the packaged behavior, but runtime compatibility requires keeping the embedded copies in chat_template.jinja and tokenizer_config.json synchronized.
