Mistral AI Drops ‘Le Chonk’: A Massive AI Model Named After a Cat Meme

Parameters serve as the adjustable numerical values that an artificial intelligence model tunes during its extensive training phase, functioning essentially as the neural connections of the system. Generally, a higher parameter count correlates with an increased capacity to learn complex patterns, handle nuanced reasoning, and process vast amounts of information, though it also drives up the computational resources and financial cost required to operate the model.

Despite its colossal 1 trillion total parameters, Mistral Large 4 employs an advanced architectural approach known as a "mixture of experts" (MoE) design. Rather than engaging the entire network for every single user query, the model dynamically routes each incoming question to a select few specialist sub-networks. As a result, only 49 billion parameters actively fire per generated answer, balancing immense foundational capability with operational efficiency. This builds upon the architecture of Mistral’s previous flagship model, Large 3, which utilized a similar routing mechanism with 41 billion active parameters out of a total 675 billion.

Guillaume Lample, Mistral AI’s chief scientist, took to social media to highlight the significance of the release. Writing on X, Lample stated that Mistral Large 4 stands firmly at the frontier of open models, describing it as by far the strongest open-weight model originating from either the United States or Europe. Open-weight models are distinct from fully closed systems because their underlying trained numbers, or weights, are made available for outside developers to download, inspect, and deploy independently. Mistral has confirmed that it will publicly release the weights for Large 4 by the end of October, enabling the broader developer community to independently test and verify the company’s performance claims.

The release of the model also brings a lighthearted community meme into official corporate lore. In June, following Mistral’s decision to rename its conversational assistant from Le Chat to Vibe, online communities on Reddit and X playfully invented a fictional, hyperbolically powerful model dubbed "Le Chaton Fat"—roughly translating to "the fat kitten." This imaginary creation was said to feature 30 trillion parameters, generate 1,000 meows per second, and effortlessly outperform competing systems like Claude Fable 5.

Mistral CEO Arthur Mensch embraced the joke at the time, replying that the system was actually called "le gros chaton," or "the big kitten" in French. The company subsequently added a cartoon cat illustration to its Vibe website. In a nod to its online fanbase, Mistral’s official launch documentation now formally refers to the new model as, "very officially," le Chonk.

Beneath the playful naming conventions lies a serious business model focused on "sovereign AI." Mistral specializes in providing models that individual countries or private enterprises can wholly own, host, and operate internally without needing to hand sensitive data over to third-party tech giants. This approach has proven highly lucrative. In August, the company secured a major partnership worth hundreds of millions of euros with HUMAIN, a state-backed entity in Saudi Arabia, specifically centered around this sovereign deployment model.

The commercial momentum continued into September when Mistral closed a massive €3 billion, equivalent to approximately $3.37 billion, Series D funding round. Led by Samsung, the financing valued the Paris-based startup at over €21 billion, roughly $23.6 billion. Company executives have noted that Mistral Large 4 serves as the very first major technical milestone on the strategic roadmap funded by this substantial capital injection.

In terms of commercial pricing, Mistral has positioned Le Chonk as a highly cost-effective alternative to established American frontier models. The system is billed based on tokens—chunks of text where each token represents roughly three-quarters of a word. Mistral charges $1.36 per million input tokens and $4.18 per million output tokens for Large 4. By comparison, Anthropic’s Claude Opus 5.5 commands $4 per million input tokens and $20 per million output tokens, while OpenAI’s GPT-6 Astra charges $10 for input and $50 for output. This pricing structure makes Mistral Large 4 roughly one-third the cost of Opus 5.5 on input and one-fifth on output, while coming in at approximately one-seventh and one-twelfth the price of GPT-6 Astra, respectively.

What the benchmarks say

When detailing the performance metrics of Mistral Large 4, the company’s official announcement predominantly benchmarks the new model against leading Chinese open-weight systems, including DeepSeek V4 Pro, Kimi K3, GLM-5.3, and Qwen3.8 Max. Proprietary American models from OpenAI and Anthropic appear only sparingly in the published comparisons, with the newest iteration of Claude, Opus 5.5, referenced exclusively within a specific cybersecurity evaluation.

Independent and standardized evaluations provide a broader picture of how Large 4 stacks up against industry peers across various computational tasks. In a blind human evaluation of coding quality conducted by Surge AI, Mistral Large 4 secured the second position out of five tested models, achieving a score of 3.74 out of 5, trailing slightly behind Claude Opus 5, which scored 4.22.

On AutomationBench, a benchmark designed to test an artificial intelligence across 657 simulated business software chores spanning finance, human resources, sales, and customer support, Large 4 attained a score of 59.9 points. The benchmark awards zero credit if a model breaks a rule or fails to complete a task’s specific goals. On Artificial Analysis’s public leaderboard for the same evaluation, Claude Sonnet 5.5 reached 71.8 points, Opus 5.5 scored 69.5 points, and Google’s Gemini 4 Argon led the board with 77.5 points.

Software engineering evaluations show similarly competitive results. On DeepSWE 1.1, a benchmark frequently utilized by developers to gauge a model’s practical coding proficiency, Mistral Large 4 achieved a score of 62. This performance places it ahead of GLM-5.3 at 61 and DeepSeek V4 Pro at 57, though it remains behind Kimi K3, which recorded a score of 68. Meanwhile, Datacurve’s internal leaderboard lists both GPT-6 Astra and Claude Opus 5 at 74 points.

With the model’s public weight release scheduled for later this month, developers and researchers worldwide will soon have unrestricted access to evaluate le Chonk’s capabilities firsthand in real-world production environments.