Mistral releases a 1-trillion-parameter model as 'open weights', and readers go straight to clause 2

A 1T-parameter MoE that Mistral says nearly ties GLM-5.3 on coding

Mistral's announcement describes a natively multimodal mixture-of-experts model with 160+ languages and three reasoning modes (low, medium, high by default). The Hugging Face repo is named Mistral-Large-4-1T-A52B, which points to roughly 52B parameters active per token. Weights ship on Hugging Face for vLLM or SGLang, with hosted access through Mistral's API, OpenRouter, NVIDIA endpoints and Microsoft Foundry.

The benchmark claims are vendor-reported. Mistral puts the model at 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3 while using 47% fewer tokens per task, and at 41% on Terminal-Bench 4.0, up 13 points from the preview. On human-preference ratings it says only Claude Opus 5.5 ranked higher. API pricing is $0.70 per million input tokens, $0.07 cached and $2.10 output. A full technical report is promised "soon", and a cyber-specialised variant goes only to vetted partners after Mistral saw some misuse.

The licence clause that the top comment quotes in full

The release comes under what Mistral calls the Mistral Ecosystem Open-Weight licence. The top thread comment pasted the restriction and drew the comparison to Redis and Terraform:

"If the Licensee or any of its affiliates operates a Model as a Service business, the Licensee must enter into a separate commercial license with Mistral before using the Model or its derivative works for any commercial purpose." — a commenter quoting the licence text

Other readers split the vocabulary rather than the facts. One pointed to the Open Source Initiative's definition of open-source AI and concluded the release fails its "use for any purpose without having to ask for permission" test (a commenter citing the OSI definition). Another argued the weights are closer to a binary than to source code.

The defence was pragmatic:

"When most people discover that certain restrictions exist on using the weights, they will neither be surprised nor disappointed, because the main thing that's being advertised - the weights are available for download - is true!" — a commenter defending the label

The worry is who can afford to keep releasing

The more durable argument was about money, not words. One commenter argued that as long as frontier training stays capital-intensive, a genuinely open-source community around frontier models is "virtually impossible" (hintymad). The same thread had another reader fearing that open models in hand today will be obsolete in a few years, with no guarantee that labs keep publishing.

For a team that wants to self-host, the clause that matters is narrow: it bites only if your product is itself hosted model access. Everyone else can read the licence once and move on.