Mistral Large 4 open weights released: 1.05T-parameter model under modified MIT license

What Mistral released

Mistral published open weights of Mistral Large 4 on Hugging Face on October 29, 2026, according to two dev.to guides posted on November 2. The checkpoints are the same 1.05T-parameter model the company serves through its API — nicknamed "Le Chonk" — rather than a distilled variant. The official repositories are not gated: the Hugging Face API reports a gated status of false for both, so any account, or even an anonymous client, can pull the files.

The model is a multimodal mixture-of-experts design with roughly 49B parameters active per token (52B counting embeddings and output layers), a 1.6B vision encoder and a 1M-token context window. Its model card lists 72 text layers and 128 routed experts, four chosen per token plus one shared expert. Mistral also says the shipped weights are not a frozen copy of the October 6 preview: the model went through roughly three additional weeks of reinforcement learning focused on coding and cybersecurity before release.

What changed since the October preview

Four things differ from the preview. Weights moved from promised-later to downloadable FP8 and NVFP4 checkpoints with no access control. Reasoning modes expanded from high or none to low, medium and high, with high as the default and low replacing none — a breaking change. The Hugging Face launch post cited by the dev.to guides notes that Transformers runs none at high, so stale settings burn tokens instead of failing loudly. API pricing is now fixed at $0.70 per million input tokens, $0.07 for cached input and $2.10 for output.

On benchmarks, Mistral reports Terminal-Bench 4.0 climbing from 28% to 41% while using about 30% fewer output tokens, and B3 Agent Security attack resistance improving from 93.3% to 97.1%. The company also claims 19.7% on Cognition's private FrontierCode 2.0 benchmark, its best open-weight score, and 48% on the Artificial Analysis Coding Agent Index, roughly level with GLM-5.3. Every figure here is vendor-reported; the dev.to authors note independent replication was still under way as of November 2.

Hardware footprint and local serving

Size, not access, is the real obstacle. The official FP8 checkpoint takes about 1,061 GB of disk and the NVFP4 version about 610 GB; community GGUF quants from ggml-org span roughly 1,115 GB (Q8_0) down to 298 GB (IQ2_XXS), and Unsloth publishes its own set. At a sustained 1 Gbps, the FP8 download runs a little over two hours.

Every official serving example assumes one node with eight datacenter GPUs and tensor parallelism of eight. The model card recommends vLLM, ships a dedicated Docker image, and requires parser flags for tool calls and reasoning; without them, thinking text leaks into content and tool calls return as plain text. SGLang serves the same topology, Transformers works via torchrun, and both engines expose OpenAI-compatible endpoints. The guides cite Hugging Face measurements of roughly 65 tokens per second on a single request, and about 480 tokens per second of aggregate throughput across 1,000 ShareGPT conversations on 8x H200, using the optional DSpark speculative decoder.

llama.cpp gained day-one support: the Q2_K quant (381 GB) fits a single Apple Silicon Ultra machine with 512 GB of unified memory. Hugging Face reports about 90 tokens per second at prefill and 15 at generation on one M3 Ultra — workable for private experiments, slow for agent loops, and 2-bit quantisation costs quality, so test against your own tasks first. Mistral recommends temperature 1.0 and top_p 0.95 at all reasoning levels. Those without hardware can call mistral-large-4 on the Mistral API, or use hosted options Mistral lists: OpenRouter, NVIDIA endpoints and NIM containers, and private Azure deployment via Microsoft Foundry.

The license

The Mistral Ecosystem Open-Weight License is MIT plus one condition. Anyone may run, distribute, fine-tune and modify the model and its derivatives, commercially or otherwise, as long as copyright and license notices are preserved. The carve-out is "Model as a Service": giving third parties access to the model or derivatives for inference or post-training requires a separate commercial agreement. Internal use, end-user products where the model powers a specific feature, relaying requests to someone else's hosted instance, and using Mistral's own products all stay outside the restriction.

Why it matters

A trillion-parameter frontier model arriving as open weights under an effectively permissive license is rare, and the only meaningful restriction targets reselling raw inference. That shifts the barrier from legal negotiation to practical infrastructure: terabyte-scale downloads and an eight-GPU node keep self-hosting a serious commitment, while the $0.70-per-million-token API remains the cheap entry point. Teams that tested the preview should audit their reasoning_effort settings before upgrading, and the headline benchmark gains deserve scepticism until independent results land.