On October 29, 2026, Mistral published the weights of Mistral Large 4, which it first released as a public preview earlier in October. The blog post says the model received "three weeks of additional reinforcement learning" since the preview. The weights are on Hugging Face, and the model is also served on Mistral's API as mistral-large-4.
What the model card says
The card describes "a 1 trillion-parameter natively multimodal Mixture-of-Experts model with 49 billion active parameters per token (52 billion including embeddings and output layers)". It uses 128 routed experts, 4 active per token, plus one shared expert. It takes text and image input, supports a 1M-token context and offers three reasoning efforts: low, medium and high, with high as the default. Mistral says it was trained on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters.
Mistral's headline results are its own or come from partners. Terminal-Bench 4.0 rose 13 points to 41 percent compared with the preview, "while reducing output tokens by 30%". The model scores 51 percent on the Artificial Analysis Cyber Index. On the Artificial Analysis Intelligence Index v5 it scores 26, and the card itself says "the leading open-weight models from China score higher on this aggregate index."
Measured: how big the checkpoints are
We summed the safetensors files listed in each Hugging Face repository on October 31. Each repo holds one set of consolidated shards, so nothing is counted twice.
- Main repository, quantization method FP8 in its config: 54 shards, 1,060.6 GB, or about 1.06 bytes per parameter.
- NVFP4 repository: 32 shards, 610.0 GB, or about 4.9 bits per parameter on average, because not every tensor is stored in 4 bits.
Calculated: which servers can hold them
The figures below are our arithmetic, not Mistral's guidance. They count weights only. The KV cache for long contexts, activations and the runtime need room on top.
- Eight H200s hold 1,128 GB in total. At the
--gpu-memory-utilization 0.85used in the card's vLLM example, that is about 959 GB, which is less than the FP8 checkpoint. - An eight-GPU B200 system with 1,440 GB holds about 1,224 GB at 0.85, leaving roughly 160 GB for cache after the FP8 weights.
- The NVFP4 checkpoint fits eight B200s with ample room. Four B200s at 0.85 give about 612 GB, which would hold the weights with almost nothing left for cache. Four 288 GB B300s leave a few hundred gigabytes free.
The card's examples all use tensor parallelism across eight GPUs, and it advises lowering --max-model-len or --max-num-seqs when memory runs short. A 1M-token context is a capability of the checkpoint, not something every deployment can afford.
The licence and the guardrails
The weights use the Mistral Ecosystem Open-Weight License, "a modified version of the MIT license", under which "operating a Model-as-a-Service business requires a separate commercial license with Mistral." Running the model for your own products is allowed. Selling access to it as a hosted model is not, without a second agreement.
On safety, Mistral reports a 96.6 percent average refusal rate on malicious cyber requests. It also writes that after watching live traffic, "we see a potential for uplift to conduct real attacks", and that it is "modifying our guardrails accordingly." A specialized cybersecurity variant goes only to vetted design partners.
What remains uncertain. The full technical report is "coming soon". The benchmarks are self-reported or partner-run. Guardrail updates reach Mistral's API, but copies of the weights already downloaded keep the behaviour they shipped with.
Analysis: Size an MoE model from its files, not from its active-parameter count. Fifty-two billion active parameters set the compute per token, while 610 GB to 1,061 GB of weights set the hardware. If Belgavi's Book Store wanted to self-host this model for help-desk triage, Tanishka would first need to decide between eight GPUs for FP8 and Blackwell parts for NVFP4. Before launching an internal API that other teams call, she should read the MaaS clause too.
Sources and evidence
Mistral released Mistral Large 4 weights on October 29 (1T-parameter MoE, 49B active per token, 1M context, three reasoning efforts); measured on Hugging Face, the FP8 checkpoint is 1,060.6 GB and the NVFP4 one 610.0 GB, so FP8 does not fit eight H200s at the card's 0.85 memory setting. The licence requires a separate commercial agreement for Model-as-a-Service use, and Mistral says it is still changing its cyber guardrails.