Mistral Large 4's Weights Are Finally Public: What 1.06 TB and a 2,632-Byte Licence Actually Get You

Twenty-three days after Mistral AI put Mistral Large 4 on sale as an API-only public preview, the checkpoint landed. The repository mistralai/Mistral-Large-4-1T-A52B went live on Hugging Face on 29 October 2026 — its metadata records a last modification at 14:11 UTC that day — carrying 54 safetensors shards, 68 files, 1,060.6 GB in bf16, and no access gate whatsoever. What arrived alongside the weights is the thing earlier coverage kept flagging as absent: a licence file, all 2,632 bytes of it, titled the Mistral Ecosystem Open-Weight License. The blanks are filled. The interesting part is what the filling says.

The timeline is worth pinning down, because three dates have been in circulation and only one of them is a delivery. Mistral announced the preview on 6 October and committed to weights "by the end of the month" without naming a day. Secondary reporting pointed at 27 October; that date passed with the repository still empty, and the pre-release write-ups — including several on this blog — correctly described a card with an "Open" badge, no licence name, and a weights tab reading "Coming soon." The repository itself had existed since 30 September as a placeholder with no files in it. It has files now.

What a reader can verify first-party, today, is the repository state: the file list, the byte counts, the licence text, and the fact that nothing asks permission before you download. What cannot yet be verified is everything that depends on someone actually running the thing at scale — whether the quantized checkpoint preserves the published scores, whether the 1M context is served rather than merely configured, and whether the deeply discounted API rate survives the month. Those are marked as open below rather than smoothed over.

What is actually inside the repository

The bf16 release is a single flat run of shards: consolidated-00001-of-00054.safetensors through -00054, fifty-two of them between 19.7 and 20.0 GB with the last two at 16.9 GB and 4.2 GB, plus a consolidated.safetensors.index.json. Total payload is 1,060.6 GB. That is a genuine 1-terabyte download before you have cached a single token of KV state, and it is the number to plan storage around rather than the "1T parameters" headline, which describes the model's arithmetic width rather than its disk footprint.

The arithmetic that matters for serving is the sparse split: roughly 1.05 trillion parameters in total, 49 billion active per token and 52 billion counting embeddings and output layers, spread across 128 routed experts with four active per token plus one shared expert. The card also claims a 1M-token context and more than 160 languages. Both are vendor ceilings until someone measures what is actually served.

• Access — ungated. Hugging Face's API reports gated: false, so there is no click-through, no approval queue, and no account requirement beyond the site's ordinary terms. The waitlist that built up around this release during October was for an announcement, not for a permission.

• Serving metadata — config.json, generation_config.json, processor_config.json, a chat_template.jinja, and both tekken.json and tokenizer.json. The processor config is there because the model takes images as well as text.

• Pipeline tag — image-text-to-text. Mistral describes the model as natively multimodal with a 1.6-billion-parameter vision encoder, text and image in, text out.

• Quantized sibling — mistralai/Mistral-Large-4-1T-A52B-NVFP4, its own repository, created 15 October, 43 files, 32 shards, 610.1 GB, also ungated. Mistral positions this one for Blackwell hardware with native NVFP4 support and says it keeps "same performance," a claim nobody outside the company has yet tested.

• Benchmark images — the repository ships exactly two charts, images/coding_index.png and images/intelligence_index.png. Both are vendor-made. They are useful as evidence of what Mistral wants you to look at, not as independent measurement, which is the distinction the next section is about.

The licence, read rather than summarised

The name on the file is the Mistral Ecosystem Open-Weight License, and its second sentence calls itself "a modified version of the MIT license." At 2,632 bytes it is short enough to read in a sitting, which is more than can be said for most open-weight terms. The grant is broad: use, copy, modify, merge, publish, distribute, sublicense, and sell copies of the model; run it, deploy it, further train it, and create derivative works. Commercial use is permitted. There is no revenue threshold, no user cap, no field-of-use restriction, and — contrary to pre-release speculation about a revenue gate — no number of any kind that a business has to stay under.

What there is instead is one condition, in clause 2. If the licensee or an affiliate "operates a Model as a Service business," it must sign a separate commercial licence with Mistral before using the model or its derivatives for any commercial purpose. That is the clause that separates open weights from an open platform, and it is narrower than it first reads. "Model as a Service" is defined as giving a third party access to the model for inference or post-training "in a manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data."

Three carve-outs follow, and they matter more than the restriction:

• End-user products — a product whose model capabilities are "solely embedded within specific features or harnesses" is outside the definition. Shipping a feature that happens to call the model is not operating a model service.

• Relaying — "mere relaying of requests to models hosted by others" is also outside it. A gateway or router that forwards to an endpoint someone else hosts is not, by this text, a Model-as-a-Service business.

• Internal use and Mistral's own products — clause 3 exempts any use that does not make the model available to third parties, and any use through Mistral's official products.

The practical reading is a three-way split. Hosting the weights inside your own organisation is unambiguously fine. Building a product around it is fine. Standing up an API that hands customers raw access to the model — the inputs, the parameters, the training — requires a commercial agreement with Mistral first. And a routing layer that relays to Mistral-hosted inference sits in the carve-out, which is precisely the sentence a gateway has to read before adding a model at all.

That distinction is worth being blunt about on our own page: Mistral Large 4 is not in OrcaRouter's catalogue as of publication, so this is not a route we can offer you today. Through Mistral's own API and a number of third-party hosting platforms it is callable now. What OrcaRouter does provide is the general alternative for the models it does carry — one API across the catalogue with provider list prices passed through at 0% markup, so a vendor price change lands the same day; automatic failover across providers; and a routing DSL for composing several models into a single call. When a model like this is unproven, failover is also the cheapest way to put it behind real traffic without betting a production path on one endpoint.

The numbers, labelled by who produced them

Every performance figure attached to this release so far comes from Mistral. No third party has published a full re-run against the released checkpoint, so what follows is vendor-reported unless the line says otherwise.

• Terminal-Bench 4.0 — 41%, up from 28% at preview. Mistral attributes the 13-point gain to three additional weeks of reinforcement learning between preview and release.

• Artificial Analysis Coding Agent Index — 48%, which Mistral describes as effectively tied with GLM-5.3 while using 47% fewer tokens per task on average. The instrument is Artificial Analysis's; the score is Mistral's report of it.

• Cyber — 51% on the Artificial Analysis Cyber Index, with a stated 96.6% average refusal rate on malicious cyber requests, and 79% on CyberGym-E2E per release coverage.

• Other vendor figures — BrowseComp 84.5%, Finch 73.3%, and B3 attack resistance 97.1%, all from Mistral's release post.

• Artificial Analysis Intelligence Index v5 — 26, which Mistral's card calls the highest score among western open-weight models. This is the number most likely to be misquoted, because the preview was reported at 38.4 on the index version standing on 6 October. Artificial Analysis rebuilt and rebased that index to v5 on 27 October, so 38.4 and 26 are readings from two different instruments. Subtracting them produces a 12-point "regression" that does not exist. Do not do it.

The one number worth carrying into a purchasing decision is not a benchmark at all: 47% fewer tokens per task on the coding agent index. If that holds under your own workloads, it changes the arithmetic far more than four points on a leaderboard, because you pay for tokens, not for rank.

What it costs to run, against what it costs to call

Self-hosting splits into two very different commitments. The bf16 checkpoint is 1,060.6 GB of weights alone, which does not fit usefully on eight 141 GB accelerators once you reserve KV cache; plan on more hardware or on the quantized path. The NVFP4 checkpoint is 610.1 GB, and Mistral's own guidance is that this version runs on four B200 or B300 GPUs. The model card's vLLM recipe asks for tensor parallelism of 8 with an FP8 KV-cache dtype, which is the configuration to benchmark against rather than a promise of throughput.

Calling it is simpler. The published rate card is $1.36 per million input tokens, $0.14 cached input, and $4.18 per million output. A launch promotion shows the same three lines halved to $0.68, $0.07 and $2.09 — the figure the product sidebar on Mistral's own open-weights post rounds to $0.70 and $2.10 — and no end date for that promotion has been published, which means the effective price of an API call today is roughly half the list rate and nobody can tell you when that stops being true.

Put those side by side and the decision mostly makes itself. For almost any team below hyperscale, four Blackwell GPUs plus the engineering time to serve a 1-trillion-parameter MoE costs more per useful token than the discounted API does. Self-hosting here is a purchase of control — data residency, licence certainty, the ability to fine-tune and ship derivatives — not a purchase of savings. That is the honest reason to download 1.06 TB.

Who should move, and who should keep waiting

Move if the missing thing was availability. A team that wanted European-trained weights it could inspect, fine-tune, or run inside its own perimeter now has them, under terms that permit exactly that, with no gate in the way. Move also if you have Blackwell capacity and a workload where the token-efficiency claim would pay for the hardware — but measure it on your own traffic first, because the 47% is Mistral's number on Mistral's chosen index.

Wait if quality was the blocker. There is still no independent evaluation of the released checkpoint, only of the preview, and the two are different artefacts separated by three weeks of RL. Wait if your fleet is H100 or H200 without native NVFP4 — the efficient path assumes Blackwell, and the bf16 path assumes a great deal of memory. And wait if the business model was reselling raw model access through your own API: that is the one use the licence clearly reserves to a commercial agreement.

What we still do not know

Four things, in order of how much they could change the picture. Whether the NVFP4 and GGUF conversions hold the published scores — Mistral says "same performance," which is a claim, not a measurement. Whether the 1M context is served or capped in practice; Artificial Analysis measured 524,288 tokens on the preview configuration, and the config file's 1,048,576 is a ceiling, not a guarantee. Whether the halved API price survives past the promo window, which determines whether calling stays cheaper than hosting. And how clause 2 will be enforced in edge cases — the phrase "meaningful control" is doing a lot of work and has no case law behind it yet.

The release itself is no longer in question. Twenty-three days of "coming soon," one missed date, and a placeholder repository became 1,060.6 GB, an ungated download, and a licence short enough that there is no excuse for not reading it. If you were waiting to decide whether Mistral Large 4 was real, the answer as of 29 October 2026 is that the weights are on disk and the terms are on paper. If you were waiting to decide whether it is good, that question has not been answered yet — and the only people who can answer it now are the ones willing to download it and run their own tests.