Mistral has published the weights of Mistral Large 4, a mixture-of-experts model with about one trillion parameters, making it the largest open-weight release yet from the French AI lab. TechCrunch and VentureBeat both reported the release on October 29, 2026, and Mistral's own announcement page confirms the details.
What is in the release
The weights are on Hugging Face under the repository name Mistral-Large-4-1T-A52B, and the model is also reachable through Mistral's API and Studio, OpenRouter, Nvidia endpoints and Microsoft Foundry, according to Mistral. Artificial Analysis lists 1.05 trillion total parameters with 52 billion active for each token, which is how a mixture-of-experts design keeps running costs below those of a dense model of the same size.
Mistral describes Large 4 as natively multimodal, covering more than 160 languages, with low, medium and high reasoning modes. The model appeared as a preview earlier in October, and this release adds the open weights and lower prices.
Pricing and licence
Mistral's API charges $0.70 per million input tokens, $0.07 for cached input and $2.10 per million output tokens. VentureBeat says that is about half the preview rates.
The licence is called the Mistral Ecosystem Open-Weight licence and is described as a modified MIT licence. According to Mistral and VentureBeat, commercial use needs a separate agreement only when the licensee or an affiliate runs a model-as-a-service business, and there is no revenue threshold. Most companies using the model internally or inside their own products would not be affected, but cloud and API resellers would.
What independent testing shows
Mistral's own benchmark claims are strong, so the independent numbers matter more. On the Artificial Analysis Intelligence Index, Large 4 scores 26, up from 25 for the preview, VentureBeat reports. That makes it the top open-weight model made outside China, but it sits six points behind DeepSeek V4.1 Flash 1017 and GLM-5.3, which both score 32.
The efficiency gains are clearer. Cost per task fell to $2.06 from $5.54, a 63 percent drop. On the Coding Agent Index Large 4 scores 48 against 49 for GLM-5.3 while using about 8.5 million tokens per task compared with 15.9 million. Its hallucination rate fell to 28 percent from 41 percent, with accuracy of 26 percent, and it scores 51 on a cyber index, third among open models.
Where the claims and the tests differ
Mistral says Large 4 reaches 19.7 percent on FrontierCode 2.0 and 41 percent on Terminal-Bench 4.0. Artificial Analysis measured 15 percent and 25 percent on the same two tests, VentureBeat notes, which is well below the vendor figures. The gap does not mean the vendor numbers are wrong, since harness and settings differ, but buyers should rely on third-party results. Mistral also reports a cyber refusal rate of 96.6 percent and resistance of 97.1 percent on a B3 attack benchmark, both self-reported.
TechCrunch reports Mistral's claim that Large 4 matches GLM-5.3 while using 47 percent fewer tokens per task. VentureBeat's reading is that the gains are incremental and mostly about efficiency and cost, not raw capability.
The sovereignty pitch
Mistral is selling openness as a security feature. Pierre Stock, the company's vice president of science, told TechCrunch the weights offer "access that can't be cut off by definition because the model is open-weighted." Chief executive Arthur Mensch said AI should "be a tool for empowerment, not dominance."
That message lands with governments and regulated companies that worry about depending on a US provider. It also fits with the wider push for national AI capacity, including Canada's decision this week to take a stake in Cohere. Running a trillion-parameter model still takes serious hardware, so for most teams the practical route will be hosted access, not self-hosting.
What it means for developers
For developers the value is a strong, cheap, auditable model with weights they can fine-tune and run in their own environment, provided they meet the licence terms. Teams that need the best raw scores will still look at the leading closed models or the two higher-ranked Chinese open models, but those carry their own compliance questions for some buyers.
What to watch next
Watch for independent coding and agent evaluations over the next few weeks, uptake on Hugging Face, and whether the US labs respond with open releases of their own. Mistral's next step will show whether it can close the six-point gap with the Chinese leaders.
Published October 31, 2026. Gadgets365 will update this article as more information becomes available.
