A frontier-scale model becomes a different kind of technology when its weights can be downloaded rather than accessed only through an API. On October 29, Mistral released the weights of Mistral Large 4, a roughly trillion-parameter model that had been available as a hosted preview since October 6. The new release includes an improved checkpoint, not just a downloadable copy of the earlier preview.

The practical change is that organizations can now inspect the model artifacts, select their own inference stack, and run or adapt the model without relying on Mistral's hosted endpoint. That is a substantial addition to the open-weight ecosystem. It is not a breakthrough in making large models run cheaply on consumer hardware. The official FP8 repository contains about 1.06 TB of files, and even aggressive community quantizations remain hundreds of gigabytes.

Two kinds of sparsity in one model

The model card describes a natively multimodal mixture-of-experts (MoE) architecture with approximately one trillion total parameters. Around 49 billion parameters participate in a token's computation, or about 52 billion when embeddings and output layers are included. Those numbers refer to computation, not the amount of model storage a deployment needs.

The text decoder has 72 layers. After an initial dense layer, each layer uses a router to select four of 128 available experts, with a shared expert also active. Rather than executing every feed-forward expert for every token, the model activates a small subset. The inactive experts still have to be stored somewhere and made accessible when a different token selects them.

There is a second form of sparsity in attention. As Hugging Face's October 29 technical walkthrough explains, Mistral combines multi-head latent attention, which compresses key/value state, with an indexed sparse-attention mechanism that selects up to 2,048 earlier positions for each query. Some key/value state is shared between layers. A learned, YaRN-scaled rotary-position scheme is part of its long-context design.

This is not a claim that Mistral invented MoE or sparse attention. Both have established precedents. What is useful is their combination in a downloadable, large-scale checkpoint with a documented architecture and working serving integrations. The design tackles two different costs: activating feed-forward parameters and attending to long histories.

A 1.6-billion-parameter Pixtral vision encoder handles image input. The model accepts text and images and produces text; it is not an audio-generation model. Mistral advertises a one-million-token context window, but that architectural limit should not be confused with the maximum context configured in every deployment.

The weights are available; the hardware barrier remains

The release has concrete deployment artifacts. Mistral provides FP8 weights and configuration files, a separate NVFP4 variant, and serving instructions for vLLM, SGLang, and Transformers. The official vLLM example uses eight-way tensor parallelism. The NVFP4 path targets NVIDIA Blackwell GPUs, so its existence does not imply equivalent acceleration on older GPUs.

The llama.cpp team's GGUF release makes the memory tradeoff unusually visible. It lists approximately 635 GB for Q4_K_M, 381 GB for Q2_K, and 298 GB for IQ2_XXS. These are model-file sizes, not complete memory budgets. Context state, runtime allocations, vision components, and other overhead can raise the actual requirement.

Hugging Face reports that a two-bit version can run on an Apple Silicon Ultra machine with 512 GB of unified memory. Its own example measured roughly 15 generated tokens per second for Q2_K on an M3 Ultra, but that is a specific reported configuration, not an independently established general speed. The more aggressive the quantization, the more important it becomes to check whether output quality survives on the intended tasks.

Long context adds another distinction. Although the model card lists one million tokens, the Hugging Face vLLM example sets --max-model-len 131072. The walkthrough also describes steep throughput degradation in one long-context llama.cpp configuration. Maximum supported context, practical serving context, and useful retrieval accuracy are three separate questions.

For a regulated organization with a substantial GPU cluster, self-hosting may be worth the operational expense. For an individual developer, the hosted API or a smaller model will often be more practical. Open weights change who can operate the model; they do not erase the economics of operating it.

Strong evidence for availability, narrower evidence for performance

There is little ambiguity about the availability claim: the checkpoint files, model card, and GGUF conversions are public. Performance claims require more care.

Mistral reports that further reinforcement learning after the October 6 preview raised its Terminal-Bench 4.0 score by 13 percentage points to 41%, while reducing output-token use by 30%. It also reports 48% on the Artificial Analysis Coding Agent Index, approximately tied with GLM-5.3 while using fewer tokens per task. Those are results presented in Mistral's release announcement, not measurements performed for this article.

Independent evaluation provides useful context. Artificial Analysis lists Mistral Large 4 at 26 on its Intelligence Index v5, above its comparison-group median of 23, and reports approximately 161 output tokens per second through Mistral's API. The index aggregates several tasks; it is not a universal measure of model quality. The speed result describes hosted serving, not the throughput a reader should expect from the downloadable weights on different hardware.

Cybersecurity is an important part of Mistral's positioning. The company reports 51% on the Artificial Analysis Cyber Index and 82% on a narrower vulnerability reproduction-and-patching test. These are different metrics, not conflicting versions of the same score. It also reports a 96.6% average refusal rate across selected harmful cyber-request tests. Mistral acknowledges a potential for the model to increase real attack capability and says it is adjusting safeguards. Neither benchmark scores nor refusal rates establish that a self-hosted deployment is safe or that a generated patch is correct.

The Reuters report on the October 29 release independently confirms the public weight release and the emphasis on defensive security. It does not constitute a reproduction of Mistral's benchmarks. Teams evaluating the model still need their own representative codebases, failure tests, security review, and measurements of total task cost.

Open-weight is not the same as unrestricted hosting

The Mistral Ecosystem Open-Weight License is based on MIT but adds a condition that changes commercial deployment choices. An entity operating a Model-as-a-Service business must obtain a separate commercial agreement before commercial use of the model or derivatives. The license defines this around giving third parties meaningful control over model inference or post-training.

The text expressly excludes embedded end-user features and mere relaying of requests from that definition, and provides exceptions for internal use and use through Mistral's official products. This is more nuanced than a blanket prohibition on commercial applications, but it is not the standard MIT license. Organizations building hosted model platforms should read the actual license rather than relying on the term open.

The release also does not make the training run reproducible. Mistral says the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in European datacenters. The weights and inference instructions are available, but the original training corpus, complete training recipe, and full technical report are not part of this release; Mistral says the report is forthcoming. Inspecting and adapting a checkpoint is not the same as being able to recreate it.

What this changes

Large 4 is consequential because it places a capable, multimodal, frontier-scale model behind a downloadable artifact rather than a single provider's API. Organizations with the hardware and expertise can inspect its architecture, run their own evaluations, customize it, and keep inference within their chosen infrastructure. Researchers can study its behavior and quantization tradeoffs directly.

The limits are equally material. A trillion-parameter checkpoint is not suddenly a laptop model because only a fraction of its experts are active per token. A million-token context specification is not a guarantee of practical million-token serving. The published benchmark results need task-specific replication, and the hosting license is a real constraint for some businesses.

Those qualifications are why the weight release matters more than the preview's leaderboard position. It expands the set of powerful models that can be independently deployed and scrutinized, while making the remaining engineering and governance costs unusually explicit.

Sources

• Mistral, "Mistral Large 4 (Le Chonk) is now open weight" (October 29, 2026)
• Mistral, Mistral Large 4 model card and serving instructions
• Mistral, official FP8 checkpoint file tree
• Mistral, official NVFP4 checkpoint
• Mistral, Mistral Ecosystem Open-Weight License
• Hugging Face, "Welcome Le Chonk: Mistral Large 4 on Hugging Face" (October 29, 2026)
• ggml-org, Mistral Large 4 GGUF quantizations and file sizes
• Artificial Analysis, Mistral Large 4 model evaluation, performance and price
• Reuters, "Mistral AI publicly releases Large 4, targets cybersecurity users" (October 29, 2026)