Mistral Large 4 Goes Open: Price vs DeepSeek, Kimi, GLM

On October 29, 2026, Mistral published the weights of Mistral Large 4, the flagship model it had put into public preview on October 6, and announced a direct API price of $0.70 per million input tokens, $0.07 for cached input and $2.10 per million output tokens, effective the same day. The weights ship under the Mistral Ecosystem Open-Weight License, which Mistral describes as a modified MIT licence with one condition attached. For teams that had already costed Large 4 from an earlier price, or had ruled out self-hosting a trillion-parameter model, the sum has changed.

Three questions follow from the release and this post takes them in order. What exactly can you do with the weights, in the licence's own words. How the new direct price compares with the open-weight models it is competing against, read from each vendor's own page rather than a reseller. And what self-hosting a model of this size actually requires, using the deployment recipes Mistral and Hugging Face published rather than a guess. We looked at the preview in our October 6 post on whether an agent team should try it; this is the follow-up now that the weights and the price are public.

Mistral's announcement is short on new capability and long on distribution. The model is the same Mistral Large 4 that entered public preview earlier in October, which Mistral describes as its flagship one-trillion-parameter, natively multimodal model, trained from scratch on European infrastructure. What is new is that the weights are downloadable from Hugging Face, the model is packaged for vLLM, SGLang and an NVIDIA NIM container, and the metered price has been set at the level above.

The documentation model card gives the specification. Two API names, mistral-large-4 and mistral-large-4-0; 1.05 trillion total parameters and 52 billion active per token in a granular mixture-of-experts design; a 1.6-billion-parameter vision encoder; a one-million-token context; text and image in, reasoning and text out. Three reasoning modes, low, medium and high, with high as the default. The card still shows a status of public preview and a release date of October 6, so the October 29 event is a weights and price change, not a version bump.

One number is unstable across Mistral's own channels and worth flagging before anyone copies it into a slide. The docs card says 52 billion active parameters and the Hugging Face repository name says A52B. The Hugging Face launch blog, published the same day, says about 49 billion active per token, with 128 experts, four routed and one shared. This post uses the docs card's 52 billion because it is the figure on Mistral's own documentation, and notes the discrepancy rather than resolving it.

The licence file in the Hugging Face repository is titled Mistral Ecosystem Open-Weight License and opens by stating that it is a modified version of the MIT licence. The grant is as broad as MIT: permission, free of charge, to use, copy, modify, merge, publish, distribute, sublicense and sell copies of the model, to run, deploy and further train it, and to create derivative works. The modification is a single numbered condition, and because the condition is what decides whether a business can use the weights at all, the sentence is reproduced here as written.

Two carve-outs follow the definition. The licence says Model as a Service does not include end-user products in which the model's capabilities are embedded within specific features or harnesses, and does not include merely relaying requests to models hosted by others. Clause 3 then exempts internal use, defined as any use that does not make the model available to third parties, and any use through Mistral's official products.

In practice that sorts readers into three groups. A company that runs the weights on its own hardware for its own staff is covered by the internal-use exemption. A company that ships a product with the model inside a feature, where customers never choose the prompts or parameters, is covered by the embedded-product carve-out. A company that sells an endpoint, a hosted fine-tuning service or a model marketplace is operating Model as a Service and needs Mistral's commercial licence first. The failure case is the middle group drifting into the third: a product that starts as a fixed feature and later exposes a system prompt, a temperature setting or a fine-tuning upload to customers has, on the licence's definition, started giving third parties meaningful control. That is the point to call Mistral. How this clause compares with the bespoke terms other open-weight vendors use is a question we covered for GLM-5.3 in our post on GLM-5.3's non-MIT licence.

Mistral's sentence is unambiguous about the new rate and silent about the old one. The news page states that starting October 29, Mistral Large 4 is priced at $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens, and the docs model card shows the same three figures. Nowhere on either page is the preview price stated, so this post does not claim a percentage cut.

What can be said with a source is this. In our October 28 capture of the OpenRouter catalogue, the route mistralai/mistral-large-4-0 listed $1.36 per million input, $0.14 cached and $4.18 per million output. That is OpenRouter's listed route price on October 28, which can reflect provider mix and margin, and it is not evidence of Mistral's own previous list price. By November 5 the same route showed $0.70 and $2.10, matching Mistral's announcement. We track those catalogue movements in our October OpenRouter price-change study.

The comparison that matters is with the other open-weight models a team would shortlist for the same work, and the honest way to do it is direct list prices only. Every figure in the table below was read from the vendor's own pricing page on November 5, 2026. DeepSeek publishes peak and off-peak rates, with off-peak at half of peak during the hours it defines; the table shows peak, the higher of the two rates; peak hours are 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday, so most US working hours fall off-peak. Kimi lists a cache-hit rate and separate cache-write charges; the table shows cache hit. Z.ai lists cached input and a storage charge that is currently free for a limited time.

The pattern is clearer on output than on input. On output, Large 4 is cheaper than every frontier-tier open model in the table and only DeepSeek's small Flash tier undercuts it. On input it sits below GLM-5.3 and DeepSeek V4 Pro and well below both Kimi models, but above V4.1 Flash. The cached-input column is where DeepSeek's pricing is most aggressive: its cache-hit rates are a few cents per million, against Mistral's $0.07, so a workload that is mostly repeated prefix, such as an agent re-reading the same repository, narrows the gap considerably.

A price table is not a quality table, and this one says nothing about which model does the work better. It also compares models of different sizes: Large 4 is a trillion-parameter model and V4.1 Flash is DeepSeek's small tier. The right reading is that Mistral has priced a flagship-sized model at the level of other vendors' mid tiers, which is the change the October 29 price makes and the reason to re-run a cost model rather than copy an old one. For the broader table across every vendor, our frontier model API price index carries the full matrix.

Mistral's own page keeps the hardware question short. It says the model can be deployed on your own infrastructure with vLLM or SGLang, points to NVIDIA's NIM container, NeMo AutoModel for customisation and TensorRT LLM for high-throughput serving, and gives no GPU count, memory figure or throughput number. The concrete recipes come from the Hugging Face launch blog, published October 29, and they are specific enough to budget from.

For production serving, Hugging Face's Inference Endpoints recipe runs the FP8 weights on eight NVIDIA H200 GPUs with tensor and expert parallelism, and its vLLM example uses the NVFP4 checkpoint, which it says requires Blackwell-generation GPUs, at tensor-parallel size eight and a 131,072-token maximum model length. The Dell Enterprise Hub listing names H200, B200, GB200 and GB300 as supported parts. The transformers example launches with eight processes per node. Eight data-centre GPUs is the baseline for the Inference Endpoints recipe, and the one-million-token context is not what those recipes run at.

For a workstation or a single machine, the blog reports that the 2-bit GGUF quantisations run on one Apple Silicon Ultra machine with 512 GB of unified memory, at roughly 90 tokens per second prefill and 15 per second generation on a Q2_K build, and that four RTX 6000 Pro cards run the NVFP4 build at about 10 tokens per second generation at 32K context. Those are Hugging Face's own measurements on its own hardware, and 2-bit quantisation is a quality trade a team should test before relying on. The failure case is reading the 512 GB figure as a laptop and the 15 tokens a second as interactive; it is neither, and the full-precision model needs a rack. We set out the same arithmetic for a smaller open-weight model in our GLM-5.2 self-hosting reality check; Large 4 runs on the same eight-H200 class of hardware in Hugging Face's recipe, with a larger total parameter count.

Mistral's news page calls Large 4 the best open-weight model for coding and backs that with a list: 19.7% on Cognition's FrontierCode 2.0, 48% on the Artificial Analysis Coding Agent Index, 41% on Terminal-Bench 4.0 described as 13 points above the preview, 51% on the Artificial Analysis Cyber Index, 84.5% on BrowseComp and 97.1% attack resistance on the B3 agent security benchmark. Every one is a figure Mistral published about its own model, and the post labels them as such.

The complication is that the Hugging Face launch blog for the same model, published the same day, reports different numbers for what look like the same tests: 28.3% on Terminal-Bench 4, 49.8% on the Coding Agent Index, 93.3% on the B3 benchmark and 82% on the Artificial Analysis Cyber Index. Some of these may be the preview model's scores, some a different harness or date, and the blog does not say. The useful conclusion is not which figure is right. It is that Mistral's news page and the Hugging Face launch blog can differ by 13 points on the same benchmark, which is a strong argument for running your own evaluation before any of these numbers reach a procurement document.

With the price, the licence and the hardware on the table, the decision is mostly mechanical. The direct API at $0.70 and $2.10 is the default for any team whose usage would not keep eight H200s busy, which is most teams. Self-hosting earns its cost only at sustained high volume or where data cannot leave the building, and then the licence's internal-use exemption applies. Reselling inference, in any form that gives customers control over prompts or parameters, needs Mistral's commercial licence before launch.

Mistral also mentions a cybersecurity-specialised variant tuned for penetration testing, vulnerability assessment and red-teaming, in preview for vetted enterprise design partners with Orange Cyberdefense named, under contractual safeguards that confine testing to owned or authorised systems. That variant is not the open-weight release and should not be budgeted as if it were. It belongs with the other gated cyber models we track in our vetting-programme ledger.

If you want help building that cost model, or standing up an evaluation that tests Large 4 against the DeepSeek and GLM tiers on your own tasks, it is part of what our AI transformation engagements cover.