Mistral released the weights of Mistral Large 4, a 1 trillion-parameter multimodal model, on October 29 under a modified MIT license. Internal use is unrestricted, and any company that operates a Model-as-a-Service business needs a separate commercial license from Mistral.
What Mistral released
According to Mistral’s release post, the model it calls Le Chonk entered public preview on October 6 and has had three more weeks of reinforcement learning since. The company says it scores 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3, the leading open-weight model, while using 47% fewer tokens per task on average. On Terminal-Bench 4.0 it reports 41%, up 13 percentage points from the preview, with 30% fewer output tokens. On Cognition’s private FrontierCode 2.0 benchmark it reports 19.7%, the top score in its comparison group.
For security work, Mistral reports 51% on the Artificial Analysis Cyber Index and an average refusal rate of 96.6% on malicious cyber requests across JailbreakBench, StrongREJECT and AgentHarm prompts. It says that is the highest it measures among open models. Booz Allen, Cisco, Cloudflare, CrowdStrike and Elastic are quoted in the post. All benchmark figures are Mistral’s own or from the evaluators it names, and we have not reproduced them.
The Hugging Face model card lists a 1M-token context window, text and image input, three reasoning levels (low, medium and high, with high as the default) and 49 billion active parameters per token, or 52 billion including embeddings and output layers. Mistral says it trained the model on 3,800 NVIDIA Grace Blackwell GPUs in its own datacenters in Europe.
The license is the part to read
The Mistral Ecosystem Open-Weight License grants the usual MIT rights to use, copy, modify, distribute and further train the model. Section 2 adds a condition. A licensee, or any of its affiliates, that operates a Model-as-a-Service business must enter a separate commercial license with Mistral before using the model or its derivatives for any commercial purpose.
The license defines the term as giving a third party access to the model or its derivatives for inference or post-training, by API or hosted endpoint, in a way that lets that party exercise meaningful control over inputs, parameters or training data. It excludes end-user products where model capabilities are solely embedded in specific features, and the mere relaying of requests to models hosted by others. Section 3 exempts internal use, defined as any use that does not make the model or derivatives available to third parties, and any use through Mistral’s official products.
What it takes to run
Hugging Face’s launch post, which comes from a distribution partner rather than Mistral, gives the first concrete hardware picture. The NVFP4 checkpoint has a smaller memory footprint and needs NVIDIA Blackwell GPUs. The vLLM example launches it with a tensor-parallel size of 8, and the Inference Endpoints recipe runs the FP8 weights across 8 NVIDIA H200 GPUs. Community 2-bit quantizations run on a single Apple Silicon Ultra machine with 512 GB of unified memory. On an M3 Ultra, Hugging Face measured prefill of about 90 tokens per second and generation of about 15 tokens per second for the Q2_K build.
Those numbers frame the gap between “open weights” and “deployable.” A trillion-parameter model is a data center workload at full precision, and the laptop-class builds trade quality and speed for footprint. Hugging Face does not publish quality figures for the quantized builds.
What it means for the buying committee
In our reading, three groups of buyers face different questions.
Enterprises that deploy the model for their own staff and systems fall under the internal-use exemption in section 3. Their decision is cost and capability: the 47% token reduction Mistral reports only matters if it holds on their workloads, so it belongs in a pilot with their own repositories and a token budget. Mistral’s own coding comparison is against GLM-5.3, so teams already running other open-weight models have a direct baseline to test against.
Systems integrators, software vendors and internal platform teams that expose the model to other companies or business units outside the corporate family need legal review before they build. The definition turns on third-party access and meaningful control over inputs, parameters or training data. A fine-tuning service or a hosted endpoint plausibly sits inside it. A product that embeds the model in a fixed feature plausibly sits outside it. Which side a given design falls on is a question for counsel and, per the license, for Mistral through its contact form.
Security leaders have a separate decision. Mistral positions the model for defenders because closed models can refuse vulnerability research, and open weights keep the usage policy under the customer’s control. That argument rests on vendors’ cyber numbers, and the comparison arrives as researchers report AI helping build exploits. Ask which refusal policy your own deployment will inherit and who owns the safeguards once the weights sit on your hardware.
The distribution path adds a vendor dependency. The weights are distributed through Hugging Face, which NVIDIA agreed to buy for $12.9 billion, and the NVFP4 build is tuned to NVIDIA Blackwell hardware. Teams budgeting for self-hosting should price the GPUs first. Teams comparing vendor pricing can set that cost beside Google’s Gemini 4 Argon introductory pricing, which ends on November 30.
Actions for the CIO and CAIO
Add the license text to the model intake checklist, and record whether each planned use is internal, embedded or hosted for third parties. Run the coding and security claims against your own code before any procurement decision, and ask Mistral in writing about any design that touches the Model-as-a-Service definition. Budget for the hardware tier the Hugging Face recipes imply, which starts at eight data center GPUs for the FP8 build.
Source: Mistral
