Paris-based Mistral AI has launched Large 4, a new artificial-intelligence model with 1 trillion parameters and a mixture-of-experts architecture designed to use only a portion of that capacity for each request.
The model was released on October 6 and has been informally branded “Le Chonk,” a reference to an online cat meme that emerged around Mistral’s chatbot earlier this year. Mistral says it plans to publish the model’s weights by the end of October, potentially allowing outside developers to download and evaluate the system themselves.
According to Decrypt, Large 4 activates 49 billion parameters for each query, while its predecessor, Large 3, activated 41 billion out of a total 675 billion. Parameters are adjustable values learned during training; in broad terms, a larger number can give a model more capacity, although operating costs also depend on how much of the system is used for each request.
A trillion-parameter model with a meme-inspired name
Large 4 uses a mixture-of-experts design. Instead of running every part of a model for every question, this architecture routes a request to selected specialist sub-networks. Mistral says the result is a model with a trillion total parameters but 49 billion active parameters per answer.
Mistral chief scientist Guillaume Lample described Large 4 on X as “at the frontier of open models” and said it was the strongest open-weight model from the United States or Europe. Open-weight models are systems whose trained numerical weights are made available for others to download and run, rather than being accessible only through a company’s online service.
The “Le Chonk” label grew out of a June joke on Reddit and X. After Mistral renamed its Le Chat assistant to Vibe, users imagined a fictional model called Le Chaton Fat, or “the fat kitten,” and attributed exaggerated specifications and performance claims to it. Mistral chief executive Arthur Mensch joined the joke, and the company later added a cartoon cat to the Vibe website. The new launch announcement now presents “le Chonk” as the model’s unofficial nickname.
Pricing and Mistral’s wider strategy
Mistral is charging $1.36 per million input tokens and $4.18 per million output tokens for Large 4. Tokens are small pieces of text used to measure how much information an AI system receives and generates. Input pricing applies to a user’s request and other material sent to the model, while output pricing covers its response.
The figures are below the prices listed in the source material for competing models. Claude Opus 5.5 is priced at $4 per million input tokens and $20 per million output tokens, while GPT-6 Astra is listed at $10 and $50. Large 4 therefore costs about one-third as much as Opus 5.5 for input and one-fifth as much for output. Compared with Astra, its input price is roughly one-seventh and its output price roughly one-twelfth.
The launch also fits Mistral’s focus on “sovereign AI”: models that governments or companies can own and operate without sending their data to outside providers. Saudi Arabia’s state-backed HUMAIN signed a deal worth hundreds of millions of euros with Mistral in August for this purpose.
Mistral raised €3 billion, or $3.37 billion, in a Series D funding round in September. The round, led by Samsung, valued the company above €21 billion, or $23.6 billion. Mistral says Large 4 is the first milestone on the roadmap supported by that financing.
How Large 4 performed in tests
Mistral’s launch materials compared Large 4 mainly with Chinese open-weight models, including DeepSeek V4 Pro, Kimi K3, GLM-5.3 and Qwen3.8 Max. The company included fewer comparisons with leading closed models such as Claude and GPT.
The results varied by test:
- In Surge AI’s blind human assessment of coding quality, Large 4 scored 3.74 out of 5 and ranked second among five models. Claude Opus 5 scored 4.22.
- On AutomationBench, which measures performance across 657 simulated finance, human-resources, sales and support tasks, Large 4 scored 59.9. Claude Sonnet 5.5 scored 71.8, Claude Opus 5.5 scored 69.5, and Gemini 4 Argon led that board with 77.5.
- On DeepSWE 1.1, a coding benchmark, Large 4 scored 62. It exceeded GLM-5.3 at 61 and DeepSeek V4 Pro at 57, but trailed Kimi K3 at 68. Datacurve’s leaderboard listed GPT-6 Astra and Claude Opus 5 at 74.
Mistral’s planned release of the weights by the end of October would give independent developers a way to examine and test the model beyond the company’s launch comparisons.
Current market picture
Moneyiar market data at approximately 10:28–10:29 on October 7, 2026, showed the following prices and changes:
- Decimal (DEL): $0.3870816323, up 0.78% over 24 hours and down 7.89% over seven days.
- BAS: $0.02251144, down 4.29% over 24 hours and 0.46% over seven days.
- ONEchain (ONE): $0.1381282093, down 5.71% over 24 hours and up 0.08% over seven days.
- Astar (ASTR): $0.0070049252, down 7.73% over 24 hours and 5.56% over seven days.
