Mistral Large 4 Review (2026): Open Weights, AA 26, $0.70/$2.10

A trillion parameters you can download, a price card that undercuts almost everyone, and a licence that stops one clause short of MIT. Here is what the weights actually cost to run and where the model really ranks.

By AI Tool Scout · Published November 1, 2026 · 19 min read

What Is Mistral Large 4?

Mistral Large 4, codenamed Le Chonk, is Mistral AI's largest model and the first trillion-parameter system the French lab has released with downloadable weights. It went into public preview on October 6, 2026 on Mistral Studio with a promise that the weights would drop at the end of the month, and it shipped on October 29, 2026 as a checkpoint, an API model, an OpenRouter listing and an NVIDIA NIM container on the same day.

The three weeks in between were not idle. Mistral describes them as additional reinforcement learning plus feedback from customers and from the security partners it had given reduced-moderation access during the preview, and the results show up as efficiency rather than a new model: cost per Intelligence Index task fell 63%, Terminal-Bench 4.0 gained 13 percentage points with 30% fewer output tokens, and the index itself moved one point, from 25 to 26. The family identity is unchanged, which matters if you evaluated the preview — more on that in What We Could Not Verify.

The specification, as published

The 49B versus 52B discrepancy is worth thirty seconds, because it looks like a spec change and is not one. The Hugging Face model card says about 1T total and 49B active; the API documentation and the release note say 52B active. The difference is counting: 49B is what fires per token inside the mixture, and 52B adds the embedding and output layers, which run on every token but sit outside the routing. Neither figure is wrong, and the gap between them is not evidence that Mistral grew the model during the preview. For planning purposes use 52B for memory bandwidth per token and 1.05T for how much silicon you need.

What the model is, in plain terms: a very wide mixture of experts that touches a small slice of itself per token. That is the same trick DeepSeek and Kimi use, at a smaller total size than Kimi K3.1 Pro's 2.8T but with a similar active budget of 104B there against Mistral's 52B. The design explains both halves of the review: it is why 1M context is affordable to serve at $2.10 an output million, and it is why the download is a terabyte.

Natively multimodal, and narrower than that sounds

Mistral calls the model natively multimodal, and the Pixtral encoder is why it can say so: images are not bolted on through a wrapper, they are part of the training mix. In practice the supported surface is text and images in, text out. There is no audio, no video, no image generation. If your use case is screenshots, scans, chart reading, diagrams, UI review or document layout, this model is aimed at you. If it needs to listen to a call, that is Mistral Voxtral 3, a different model from the same lab with a very different price structure.

The vendor's boldest multimodal claim is visual grounding, where the preview post said the model surpasses even frontier closed models. We have not reproduced that and it is not in the third-party numbers below, so treat it as the most interesting unverified claim in the release rather than as a feature.

The Weights and the Licence

Two things decide whether "open weight" is meaningful to you, and they are separate questions: what can I actually download, and what am I allowed to do with it. Mistral answers the first generously and the second almost generously.

What is on the shelf

• FP8 — the official high-fidelity checkpoint. About 1,050 GB of weights, the version you serve if you have a serious node and want the reference behaviour.
• NVFP4 — the official 4-bit FP4 build, around 525 GB, for Blackwell-class hardware and for fitting the model into half the memory. Expect a small quality delta on the hardest tasks; that is the normal cost of FP4 and Mistral ships it as a first-class variant rather than a community hack.
• GGUF — built by Unsloth and the llama.cpp team. These exist so people can experiment and evaluate on accessible hardware. They are not a production serving path for a 1.05T model at any real traffic, and anyone who tells you they are has not looked at the memory bandwidth.
• Serving stacks — vLLM and SGLang supported the model on release day, which is unusually fast for a new architecture of this size and is a real signal about how the release was prepared. The vLLM write-up documents AMD MI350 with ROCm as a supported target, which is the first time a model of this class has had a credible non-NVIDIA path on day one.
• Hosted routes — the Mistral API and Studio, OpenRouter, NVIDIA-hosted endpoints with an optimised NIM container, and TensorRT-LLM for throughput. Mistral Forge handles enterprise customisation and Vibe Code gives you a coding agent pointed at a repository.

The absent item is the technical report. Mistral says a full one is coming soon. Until it lands, there is no published account of the training data mix, the number of experts, the routing scheme, the tokenizer, or how the 1M context was trained — and those are the details a serious self-hosting or fine-tuning decision needs. A checkpoint without a paper is a very large black box you are permitted to own.

The licence, clause by clause

The weights are under the Mistral Ecosystem Open-Weight License, which Mistral accurately describes as a modified MIT licence. Read the sentence from the release slowly, because every clause is doing work: the model is "free to use, modify, fine-tune, and commercialize, with a single condition: businesses that operate a Model-as-a-Service offering must obtain a separate commercial license from Mistral before using the model commercially."

That last row is the one that catches people. Several open-weight licences draw a line at annual revenue or monthly active users, so a small SaaS can operate under the community terms and only the giants need to negotiate. This licence does not draw that line. The test is what your business is, not how big it is: if you operate a model-as-a-service, you need an agreement, whether you are a hyperscaler or two people with a rented node. If you build an application whose value is the workflow around the model — an accounting tool, a document reviewer, a support agent — you are inside the grant and you owe Mistral nothing.

Compare that with the rest of the market and it lands in the permissive middle: looser than a research-only or non-commercial licence, tighter than Apache 2.0, and roughly where Meta's community licence sits in spirit while being considerably shorter. For a European enterprise with a procurement review, "modified MIT with one clearly stated carve-out" is an easy paragraph to approve. For a founder planning an inference business, it is a phone call.

Pricing: $0.70 / $0.07 / $2.10

Mistral halved its own preview prices at the release, which is rare enough to be worth stating plainly: the preview charged $1.36 input, $0.14 cached and $4.18 output, and the released model charges $0.70, $0.07 and $2.10. The model did not get smaller; the serving got better and Mistral passed it on.

The worked example: 500M input, 80% cached, 50M output

A month of real work for a mid-sized team: 500 million input tokens, of which 80% is a pinned codebase, contract set or knowledge base served from cache, plus 50 million output tokens including reasoning.

Two things follow from those three numbers. First, the output leg is now the biggest line item — $105 of the $203 — so reasoning modes and verbosity control matter more than input compression, which is the opposite of how people usually optimise. Second, caching is not a nice-to-have at this price: 55% of the bill is a structural decision, not a rounding error, and any workload with a stable preamble should be engineered for cache hits before it is engineered for anything else.

Per task, where the price advantage narrows

Artificial Analysis measures the released checkpoint at $2.06 per Intelligence Index task, down from $5.54 for the preview — a 63% reduction that is the single best thing about this release for anyone with a budget. Put against the field, though, the ordering flips:

Figures from Artificial Analysis's October 29, 2026 release assessment and the model page, which also shows $2.06. These are Intelligence Index tasks, not Coding Agent Index tasks, and they are measured rather than projected from token prices.

A cheap token price and a cheap task are different things, and Mistral Large 4 is the clearest example on this site this month: $0.70 per million input looks like the cheapest serious card in the market, and $2.06 per finished task is 3.4× what DeepSeek V4.1 Flash charges for the same kind of work. The mechanism is token metabolism — the model talks more to reach an answer, so a lower per-token rate does not survive contact with a workload. The one place the efficiency claim holds is coding agents, where Mistral spends about 8.5 million tokens per task against roughly 15.9 million for GLM-5.3 at a near-identical score. Pick your benchmark carefully: on reasoning-heavy tasks this model is expensive per result, on agentic coding it is unusually thrifty.

The Three Reasoning Modes, and What They Cost

Mistral exposes reasoning effort as a request parameter with three settings, and the mapping is documented rather than implied:

Two operational notes that the documentation does not spell out. High being the default is a trap for prototypes: a developer who wires up the API and measures cost will measure the most expensive mode, conclude the model is unaffordable, and move on — when medium on a classification task might be a fifth of the bill for the same accuracy. And the modes are not a monotonic dial for every task: on long agentic coding runs, low can fail a task that high would have finished, and a failed task costs you the whole run plus the retry, which is usually worse than the tokens you saved. The right pattern is per-endpoint, measured: high for the coding agent, medium for summarisation, low for routing and classification.

The efficiency story sits under all of this. Mistral reports 30% fewer output tokens on Terminal-Bench 4.0 than the preview while gaining 13 points, and Artificial Analysis still counts about 8.5 million tokens per coding task. Both can be true: the model got thriftier and it is still a heavy consumer. Budget output tokens, not input tokens.

What the Third Party Found

Artificial Analysis published an assessment of the released checkpoint on October 29, 2026, running its own harness. This is the only independent capability evidence in the review, and it is smaller than the launch post implies.

The honest reading of 26: it is a middle-of-the-table score, and "best outside China" is a geographic qualifier doing a lot of work. It is genuinely the strongest open-weight model from Europe or the United States, which is a real achievement for a European lab and a real selling point for anyone with procurement rules that exclude Chinese providers. It is not the strongest open-weight model, and at 32 the two Chinese leaders sit six points clear — a gap roughly the size of the one between this model and models several tiers below it.

The Coding Agent Index at 48% is where the release looks best, because the score is within a point of the leader while spending about 47% fewer tokens per task. Mistral's framing — "effectively tied with the leading open-weight model while using 47% fewer tokens per task" — is fair on the first half and more than fair on the second. If you are running a coding agent at volume, tokens per task is the metric that shows up on the invoice, and this model's per-task economics in that one workflow are genuinely competitive even though its per-token economics are not.

What Mistral reports about itself

Everything below is vendor-reported: measured by Mistral, or by a partner on a harness Mistral does not control. None of it can be reproduced from outside today.

The pattern worth noticing is which claims are checkable and which are not. Terminal-Bench 4.0 is a public benchmark with a public harness, so 41% can be argued with even if this run cannot be replayed. FrontierCode 2.0 is private by design — Cognition keeps the tasks secret to prevent contamination — so 19.7% is a number you either trust or do not. Cybench at 93% is checkable in principle but measured on 40 questions, which is a sample size that supports "strong" and not "best". The 82% figure on real vulnerabilities has no published test set at all.

None of this makes the numbers false. It makes them claims about capability rather than measurements of it, and a buyer should treat the difference the way they would treat any single-source performance claim: useful for shortlisting, useless for a contract.

Pros & Cons

5 Best Mistral Large 4 Alternatives

The question is not "is Mistral Large 4 good" — it is "is it the right model for this workload", and the answer differs sharply between reasoning-heavy work and agentic coding. These five cover the cases where it is not.

Comparison Table

AA v5 scores are on Artificial Analysis Intelligence Index v5 as reported in the October 29, 2026 assessment and the models' own pages, with competitors at maximum reasoning effort unless noted. GLM-5.3 at 32 and Meta Muse Spark Open at 25 are discussed in the text rather than tabulated because we have not verified their rate cards. A model with no published v5 figure is not a model that scored zero. Prices in US dollars, checked November 1, 2026; Mistral's preview rates were $1.36 / $0.14 / $4.18 and no longer apply.

Self-Hosting Reality Check

"Open weight" and "cheap" are different claims, and this is the section where they come apart. The arithmetic below is order-of-magnitude — it ignores attention memory, KV cache, activations and parallelism overhead, all of which only make the number bigger — and it is still enough to settle the question.

Now price it. A rented 8× H200 node at a list rate of roughly $3 per GPU-hour is about $576 a day, or $17,280 a month, before storage, redundancy, networking or an engineer. Our modelled API month — 500M input, 80% cached, 50M output — is $203. That is a ratio of about 85 to 1. To break even on the hardware you would need to be pushing on the order of 40 billion input tokens a month, which is not a mid-sized team's workload; it is an inference provider's.

Three caveats keep this honest, because the comparison is not purely financial:

• Data sovereignty is worth money. If your regulator, your customer contract or your own risk appetite says the data cannot leave your premises, the API is not an option at any price, and $17,000 a month is the cost of a requirement rather than a bad deal. This is Mistral's actual market: European enterprises and public institutions buying jurisdiction, not savings.
• Fine-tuning changes the maths. A checkpoint you can adapt on your own data is a different product from an API you can prompt. If the value is in the fine-tune, the rental is R&D cost, not serving cost.
• Utilisation is the whole game. A node running at 15% utilisation is an expensive way to serve a little traffic; a node running at 85% on a high-volume batch job can undercut any token price. The break-even above assumes continuous load, and most teams do not have it.

The practical recommendation: start on the API at $0.70, measure your actual token volume for a month, and only then price a node. The download is a decision about control, and the invoice should be presented to whoever owns the compliance requirement, not to whoever owns the inference budget.

What We Could Not Verify

Four things a buyer needs are missing, and they are missing in ways that matter more than the usual gaps.

One methodological note on the index comparison, because it changes how you read every score on this page. Artificial Analysis moved from v4.3.2 to v5 on October 27, 2026, and the new version re-scored the entire field downwards — Mistral Large 4's preview went from 38 on the old index to 25 on the new one, and other models lost between 9 and 15 points in the same change. Any comparison between a v5 score in this review and a v4.3.2 score in an October review on this site is invalid. Six points between Mistral and DeepSeek on v5 is a real gap; the 38 that circulated during the preview is a different index and a different model checkpoint.

Frequently Asked Questions