Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableX: @minchoi→ original

Mistral Releases Weights for Le Chonk, a 1-Trillion-Parameter Model That Ranks #1 Among Open Models on FrontierCode

Mistral has published the weights of Le Chonk, an open 1-trillion-parameter model that can run on your own servers without sending code to an external API. It uses a Mixture of Experts (MoE) design: like a consulting firm with a thousand specialists, where only the relevant ones are called in for each question, so just 52 billion parameters are active for each token. On Cognition's FrontierCode benchmark, it ranks #1 among open models. Teams with proprietary code should test it on their own tasks right away.

AI-processed from X: @minchoi; edited by Hamidun News
Mistral Releases Weights for Le Chonk, a 1-Trillion-Parameter Model That Ranks #1 Among Open Models on FrontierCode
Source: X: @minchoi. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceX: @minchoi ↗
Editor-in-Chief
Key metric #11 trillion parameters in total, with roughly 52 billion active per token
Key metric #2An NVFP4 version for Nvidia GPUs, with Nvidia support from day one
Key metric #3Weights are already on Hugging Face; the post does not mention the license or the test figures

Picture a typical situation: you are building a coding assistant for a bank or a defense contractor. The best coding models are only available through someone else's API, and sending source code there is not allowed under security rules. Open models are usually weaker, so you have to choose between quality and control. The news from Mistral is about how that choice may become less rigid: the company has published the weights of Le Chonk, a 1-trillion-parameter model. According to a post by Min Choi on X, it is the best open model on FrontierCode, a benchmark from Cognition, and Nvidia supported it from day one.

What Was Released

Le Chonk uses a Mixture of Experts (MoE) architecture. It has 1 trillion parameters in total, but only about 52 billion are used to process each token. The weights are already on Hugging Face. There is also a version in NVFP4 format, a 4-bit number representation that Nvidia promotes for its GPUs. The fewer bits per parameter, the less memory the weights take up and the faster inference runs.

Quality was measured on FrontierCode, a benchmark from Cognition, the company that makes the coding agent Devin. The post says Le Chonk ranks first among open models. The benchmark figures themselves are not given in the source, so this data cannot be used to compare the model with closed competitors.

Making Sense of "1 Trillion, but 52 Billion Active"

Consider a consulting firm with a thousand specialists. For each client question, the firm does not bring in everyone. A manager-router selects several suitable experts. The bill for the work looks like that of a small team, while the firm's knowledge is that of a large one. MoE works the same way: the model contains many expert blocks, but only a subset is activated for each token.

In practice, generation speed and cost are closer to a model with a little over 50 billion parameters than to a dense model with a trillion. There is a caveat, though. All the weights still have to be held in memory, because it is not known in advance which experts the next token will need. By our rough estimate, at 4 bits per parameter, a trillion parameters take on the order of 500 GB for the weights alone, not counting memory for the context. That is why the NVFP4 version is not a bonus here but a requirement for running the model on a reasonable number of accelerators.

What to Do With This

For teams building coding agents or internal assistants, the news matters for three reasons.

  • An alternative to closed APIs. If the FrontierCode result holds up on your own tasks, there is an open model you can deploy in-house. This matters when code cannot be sent to external services.
  • Readiness for Nvidia. Support from day one means you do not have to wait weeks for optimized kernels and launch recipes. For production, this often matters more than a gap of a couple of benchmark points.
  • Inference economics. Fewer active parameters mean less computation per token. Under heavy load, this is what determines cost.

What Is Still Unknown

The post is very short, and many details are missing. It does not specify the license (which determines commercial use), the context window size, the hardware requirements for specific configurations, or the benchmark numbers themselves. It is also unclear how well FrontierCode reflects real-world tasks: any public benchmark should be supplemented with a set of tasks from your own repository.

A sensible plan is simple. Read the model card on Hugging Face, check the license, run a few dozen of your real tasks (bug fixes, code review, test generation), and compare the total cost of ownership with what you currently pay for API access. If the model really leads among open models and running it fits within your GPU budget, it is a strong candidate as the foundation for a self-hosted coding agent.

Story

Key Questions Answered

What Was Released?

Le Chonk uses a Mixture of Experts (MoE) architecture. It has 1 trillion parameters in total, but only about 52 billion are used to process each token. The weights are already on Hugging Face. There is also a version in NVFP4 format, a 4-bit number representation that Nvidia promotes for its GPUs. The fewer bits per parameter, the less memory the weights take up and the faster inference runs. Quality was measured on FrontierCode, a benchmark from Cognition, the company that makes the coding agent Devin.

What are the key details on Making Sense of "1 Trillion, but 52 Billion Active"?

Consider a consulting firm with a thousand specialists. For each client question, the firm does not bring in everyone. A manager-router selects several suitable experts. The bill for the work looks like that of a small team, while the firm's knowledge is that of a large one. MoE works the same way: the model contains many expert blocks, but only a subset is activated for each token.

What to Do With This?

For teams building coding agents or internal assistants, the news matters for three reasons. An alternative to closed APIs. If the FrontierCode result holds up on your own tasks, there is an open model you can deploy in-house. This matters when code cannot be sent to external services. Readiness for Nvidia. Support from day one means you do not have to wait weeks for optimized kernels and launch recipes. For production, this often matters more than a gap of a couple of benchmark points. Inference economics.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…