Summary
- Mistral introduced Large 4 on October 6, a powerful AI model featuring 1 trillion parameters, which activates 49 billion parameters per query, priced at $1.36 per million input tokens and $4.18 per million output tokens.
- The model's playful name "Le Chonk" is a nod to the "Le Chaton Fat" meme from June, with plans to make the model's weights available by the end of October.
- According to launch metrics on Finance Agent v2, Large 4 scored 54.7, surpassing GPT-6 Astra at 53.5 but lagging behind Claude Opus 5.5, which scored 58.6.
Paris-based Mistral AI unveiled its latest development, Mistral Large 4, on Tuesday, which boasts an impressive 1 trillion parameters. This type of model is similar to those behind popular chatbots like ChatGPT and Claude, where parameters are the adjustable figures that enhance learning capabilities, contributing to operational costs.
Guillaume Lample, Mistral AI’s Chief Scientist, remarked on X, “[Mistral Large 4] is at the frontier of open models, and by far the strongest open-weight model from the US or Europe.”
This model employs a "mixture of experts" architecture, activating only a subset of its parameters for each query—specifically, 49 billion—while the previous model, Large 3, activated 41 billion from a total of 675 billion parameters.
The moniker “Le Chonk” has an entertaining origin. Following Mistral's rebranding of its Le Chat assistant to Vibe, fans on social media created a fictional model called Le Chaton Fat, or "the fat kitten," which humorously claimed to have 30 trillion parameters and exaggerated performance metrics.
The French Government needs to STOP Le Chaton Fat before it's too late. This is the FAT takeoff we've been warned against for years! pic.twitter.com/VoaBPJmfBT
— GLIF (@heyglif) June 15, 2026
Mistral's CEO, Arthur Mensch, humorously responded to the meme, clarifying that the model was actually named "le gros chaton," which translates to "the big kitten." Subsequently, a cartoon cat was added to the Vibe website, and the model was officially dubbed "le Chonk."
Mistral offers "sovereign AI" solutions, which allow countries or companies to operate models independently without sharing their data with third parties. Notably, Saudi Arabia's state-funded HUMAIN recently entered into a multi-million euro agreement with Mistral for such services.
In September, Mistral completed a €3 billion ($3.37 billion) Series D funding round, achieving a valuation exceeding €21 billion ($23.6 billion), with Samsung leading the investment. Mistral notes that Large 4 marks a significant milestone in the roadmap funded by this investment.
BitcoinBTC · USD$85,712+3.17%24H7D1M1YYTDSep 29Oct 1Oct 3Oct 4Oct 6$86.8k$85.6k$84.3k$83.0k24h HighHigh$86,64824h LowLow$85,122VolVol$1.1BMarket projectionsOdds by MyriadThis weekBelow $86,000Below $86k55% chance→Buy Bitcoin with USDTPowered by Jupiter$50$100$500BuyPrice data by CoinGeckoCoinGeckoMore Bitcoin news and projections →Le Chonk's pricing structure includes $1.36 per million input tokens and $4.18 per million output tokens—units of text that AI companies typically charge for, with each token being roughly three-quarters of a word. For comparison, Claude Opus 5.5 charges $4 and $20, while GPT-6 Astra charges $10 and $50.
This pricing positions Large 4 as significantly more affordable than Opus 5.5, costing about one-third as much for input and one-fifth for output. Compared to Astra, it is approximately one-seventh and one-twelfth the cost, respectively.
Benchmark Performance
Mistral's release primarily compares Large 4 against various Chinese open-weight models, including DeepSeek V4 Pro, Kimi K3, GLM-5.3, and Qwen3.8 Max, with "open-weight" indicating that these models are accessible for anyone to download and utilize. Comparisons with Claude and GPT are limited, with the latest Claude, Opus 5.5, appearing only in a cybersecurity context.
Surge AI conducted a blind evaluation of coding quality, wherein Mistral Large 4 secured the second position among five models, earning a score of 3.74 out of 5, trailing Claude Opus 5's 4.22.
AutomationBench assessed Large 4 against 657 simulated business tasks across various domains, awarding it 59.9 points. In comparison, Claude Sonnet 5.5 achieved 71.8, Opus 5.5 scored 69.5, and Gemini 4 Argon led with 77.5.
On the DeepSWE 1.1 benchmark, which evaluates coding proficiency, Large 4 registered a score of 62. This surpassed GLM-5.3's score of 61 and DeepSeek V4 Pro's 57, but fell short of Kimi K3's 68. According to Datacurve's leaderboard, GPT-6 Astra and Claude Opus 5 scored 74.
Mistral plans to release the weights for Large 4 by the end of October, allowing external developers to download the model and verify its performance claims independently.