Mistral has released the full weights of Mistral Large 4 — nicknamed "Le Chonk" — making a roughly 1-trillion-parameter flagship model freely downloadable for anyone to run on their own infrastructure. The weights were published on Hugging Face on October 29, 2026, three weeks after the model first appeared as an API-only preview on October 6, and the release dominated AI news coverage on November 8–9, 2026.
It is, by Mistral's account, the first trillion-parameter model released with open weights by any lab, and the largest open-weight model a European lab has ever released. That scale is the headline, but the licence, the benchmarks and the economics are what will decide whether the release matters beyond the announcement.
What Is Inside: A Trillion Parameters, 49 Billion at a Time
Large 4 is a mixture-of-experts (MoE) model. Instead of activating every parameter for every token, an MoE model routes each token to a small subset of specialist sub-networks, or "experts." Large 4 has about 1.05 trillion total parameters, but only around 49 billion active parameters per token — 52 billion if embeddings and output layers are counted. The router picks 4 of 128 routed experts plus one shared expert per token.
That architecture is the reason a trillion-parameter model is downloadable at all: the full weights are roughly a terabyte at FP8 precision, which is large but within reach of serious self-hosting setups using tools like vLLM or SGLang, or as an optimised NVIDIA NIM container. The model is natively multimodal, accepting text and images, and Mistral says it was trained from scratch on European infrastructure. The reported context window is 1 million tokens, although that figure has not been confirmed in all sources.
The Licence: Free, With One Condition
Large 4 is released under the Mistral Ecosystem Open-Weight License, which Mistral describes as a modified version of the MIT licence. Users may use, modify, fine-tune, distribute and sell products built on the model, free of charge, provided Mistral's copyright and permission notice travels with every copy.
There is one condition, and it is the commercially important one: a business that runs a Model-as-a-Service offering — selling access to the model itself as a hosted service — needs a separate commercial licence from Mistral first. There is no revenue threshold; the condition applies regardless of size. In effect, Mistral is giving away the model while reserving the right to charge the companies that would resell it as a service. For a startup building a product on top of Large 4, the licence is effectively free. For a cloud provider that wants to host it for customers, it is a negotiation.
The Benchmarks: Strong, and Mostly Mistral's Own Claims
Mistral reports that Large 4 leads open models on coding, and the numbers are striking. On the Artificial Analysis Coding Agent Index it scores 48 percent, which Mistral describes as effectively tied with GLM-5.3 while using 47 percent fewer tokens per task on average — 8.5 million tokens per task against 15.9 million for GLM-5.3. On Terminal-Bench 4.0, a test of real coding-agent work, it scores 41 percent, up 13 points from the preview, and on Cognition's FrontierCode 2.0 it leads at 19.7 percent. In Surge Code's blind pairwise code evaluations, graders preferred it over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, ranking behind only Claude Opus 5.5 among the models tested.
Two caveats matter. First, most of these figures are Mistral's own claims, and independent evaluations have not yet fully caught up. Second, Artificial Analysis's own independent ranking puts Large 4 second on the Coding Agent Index rather than tied first — 49 for OpenCode with GLM-5.3 at maximum reasoning against 48 for Mistral Vibe with Large 4 — and scores it 26 on Intelligence Index v5, the highest-scoring open-weight model developed outside China, but six points behind DeepSeek V4.1 Flash and GLM-5.3 at 32. The model also scores 51 percent on the Artificial Analysis Cyber Index.
Hosted API pricing is $0.70 per million input tokens, $0.07 for cached input and $2.10 per million output tokens.
Why It Matters for Founders and Investors
Open-weight releases of this scale change the economics of building with AI. When a frontier-class coding model can be downloaded and run on private infrastructure, startups no longer have to choose between paying a closed provider's API prices and settling for a weaker open model. The token-efficiency claim is the commercially interesting part: if Large 4 really does comparable coding work with 47 percent fewer tokens, the cost of running agentic coding workflows falls sharply, which is exactly the kind of cost curve that turns an experimental tool into a production system.
The licence condition is the other half of the story. Mistral is betting that the value of a model increasingly sits in the hosting layer, not the weights. For investors, that is a live experiment in open-source business models: give away the weights, charge the resellers, and see whether the ecosystem that forms around the model is worth more than the API revenue you gave up.
Angeled.com connects founders with investors.
TechnoGalaxy, "Mistral Large 4 goes open weight: a trillion-parameter model anyone can download, with one licence condition," November 5, 2026
DepthGrid, "Mistral Large 4 Goes Open Weight: 1T MoE, Modified MIT License," November 9, 2026
Read Next
OpenAI Releases Codex Security as Open-Source CLI and SDK
Databricks Unveils Always-On Genie, an AI Agent That Stops Waiting to Be Asked
DeepMind Launches AlphaGenome 2, a DNA Model That Reads Genomes Cell by Cell
ElevenLabs Unveils 13 New Products and Upgrades at Its New York Summit
Google Launches Gemini Decision 1, a Model That Answers Instead of Writing
Runway Open-Sources Praxis-1, Its Robot Model That Learned by Watching Video



