Mistral Large 4 Open Weights: 5 Checks Before You Trust the Hype
This is a preview, not the guide. Mistral Large 4's open weights have not landed yet. What follows is everything verified about the model so far, plus the five checks that will decide whether the release deserves its hype the day the weights actually appear. The full hands-on guide follows once the weights and the licence are genuinely public — that was the plan, and this preview does not change it.
Here is the honest starting position. Mistral AI announced Mistral Large 4 — nicknamed "Le Chonk" — on October 6, 2026, at a launch event in Abu Dhabi, with an API preview live the same day. The company promised the full open weights by the end of October. Reuters reported a public release on October 27. Mistral's own Hugging Face page now estimates October 31. Neither date is Mistral-confirmed as a fixed day, and the exact weights date is, as one careful analysis put it, not fixed. So if you see anyone promising you the weights on a specific date, they are quoting a report, not the company.
That uncertainty is not a footnote. It is the shape of this whole release: a genuinely large, genuinely interesting model, wrapped in documentation that disagrees with itself on details, with its most important document — the licence — still unpublished. This preview sorts what is solid from what is soft, so that when the weights do appear, you know exactly what to verify before you build on them.
What Large 4 Actually Is
Mistral Large 4 is a granular mixture-of-experts system: most of its 1.05 trillion parameters sit idle for any given token, with only a fraction — roughly 49 to 52 billion — active at once. That sparsity is the whole point of the design: a huge model that costs a fraction of its size to run. It takes text and images in and produces text out, with a 1.6-billion-parameter vision encoder reported in launch coverage, and Mistral says it was trained across more than 160 languages, including all official EU languages.
The training story is part of the pitch, and it deserves both attention and scepticism. Mistral built the model from scratch in roughly two months on between 3,800 and 4,000 Nvidia Grace Blackwell GPUs in its own European datacenters — no US cloud rental, no dependency on someone else's infrastructure. For a continent that has spent years writing AI regulation while buying American compute, that detail matters more than any benchmark. It is also, as one analysis noted, suspiciously fast: training a trillion-parameter model from scratch in two months is either a sign of extraordinary engineering or a sign that "from scratch" is doing some heavy lifting in the marketing copy. The open weights, when they land, will let the world check the work — which is exactly why the weights matter more than the announcement.
The commercial picture around the model is substantial. Mistral raised €3 billion in a Samsung-led Series D on September 8, at a post-money valuation above €21 billion, and says it supports more than 125 enterprises, including Airbus, ASML, and HSBC. CEO Arthur Mensch expects annual recurring revenue above $1 billion this year. This is not a research demo from a small lab. It is Europe's most serious attempt yet to stop being a spectator in the frontier AI race — and the weights release is the moment that attempt gets independently tested.
One practical note before the checks: Large 4 is too large to run on a desktop or laptop, and may be impractical even for some universities. Until the weights land, the preview is served on Mistral's own infrastructure, and the only way to touch the model today is Mistral's API. If your plan involves self-hosting, your real decision starts the day the checkpoint appears, not before.
Check 1: The Date and the Download — Where Exactly Do the Weights Appear?
The first thing to verify is also the simplest: did the weights actually appear, and where? Reuters reported October 27. Mistral's Hugging Face page now estimates October 31. Mistral's own wording is "by the end of the month." Those three statements are not the same, and the gap between them is the first test of how this release will be handled.
When the day comes, check Mistral's Hugging Face page for a Large 4 checkpoint — that is where the release would be verifiable. As of launch, the page listed no Large 4 checkpoint at all, which is why the "open-weight model" label has, until now, been a calendar entry rather than a downloadable fact. An API preview is not open weights, however good the API is. The distinction matters because everything else in this preview — the licence, the benchmarks, the safety claims — can only be independently tested once the weights exist in the wild.
Also check the size and the format. A trillion-parameter model does not download casually: check the checkpoint format, the quantization options, and the hardware requirements Mistral publishes alongside it. If the release ships only in a form that requires a datacenter to run, that is a fact about who the model is really for, whatever the marketing says.
Check 2: The Licence — Read the Text Before You Build Anything
This is the check that matters most, and the one most coverage will skip. Mistral's previous flagship, Large 3, shipped under Apache 2.0 — a permissive, well-understood licence. Large 4's weights are expected under a custom Mistral licence, and as of this writing, no licence text has been published. The docs card's licence badge reads only "Open." "Open" is not a licence. It is a vibe.
Why this matters depends on who you are. Recent Mistral releases show how much licence text varies: Devstral 2, Mistral Medium 3.5, and the Timestral family ship under a Modified MIT licence whose text grants standard MIT rights and then withholds them from any company whose global consolidated revenue exceeded $20 million in the preceding month, unless Mistral grants a separate commercial licence — permissive for individuals, startups, and research, closer to a restrictive research licence for large enterprises. Voxtral TTS ships under CC BY-NC 4.0, which bars commercial use outright. Large 4 could land anywhere on that spectrum, and you cannot know until you read it.
When the licence appears, check four things in order: whether commercial use is permitted at all; whether there is a revenue or company-size threshold that changes the terms; whether redistribution of fine-tunes is permitted, and under what terms; and whether the licence can change after you have built on the model. A licence that permits research but restricts commercial use is a research artefact, not a business foundation. A licence with a revenue threshold is free for startups and a negotiation for enterprises. Until the text is public, anyone telling you Large 4 is "open source" is guessing — and the guide that follows this preview will quote the licence text itself, not the badge.
Check 3: The Benchmarks — Vendor Claims vs Independent Tests
Mistral's headline numbers are vendor-reported, and the honest way to read them is to separate what they show from what they hide.
Coding is the weak spot, and Mistral admits it. The company claims 61.7% on DeepSWE v1.1 for coding. On the same board, GLM-5.3 and Kimi K3 sit around 69%, and the leading closed models around 74%. Mistral's own coverage concedes it trails in coding, and critics add that it is late and costly per task. That gap — roughly 12 points behind the best open-weight rivals on coding — matters more than any single headline number, because coding is where most developers will actually use the model. Our Kimi K3.1 Pro guide shows what the open-weight competition looks like at the top of that board, and it is a useful baseline for judging Large 4's claims.
Cyber is the strong spot — and the one that needs replication most. Mistral claims 82% on a cyber reproduce-and-patch test and 93% on Cybench. The 82% figure already has Artificial Analysis's confirmation, according to one analysis, and the Cyber Index is expected to show whether the released weights land in the top three open models. But a Cybench replication by a party other than Mistral would settle whether the 93% holds, and until third parties test the released weights, all performance claims should be treated as marketing with good formatting. The same discipline applies here as to every benchmark: our ARC Prize Astra guide showed how a single number can mean two different things depending on the harness — and Large 4's numbers deserve the same "who ran it, under what conditions" treatment.
The documentation discrepancies are a warning light, not a verdict. The active-parameter figure is reported as both 49 billion and 52 billion. The context window is reported as both 524,000 tokens and 1 million tokens. The API pricing in the launch post ($1.36 per million input tokens, $4.18 per million output) reportedly appears at half those rates on the model page. Each discrepancy is small. Together, they suggest preview documentation assembled quickly — which is understandable three weeks before a weights release, and is also exactly why the release-day documents, not the launch-week ones, are the ones to trust. When the weights land, check the model card's spec table against every launch claim, and note every difference. The differences are data.
Check 4: The Safety Story — Red-Teaming Promises vs Published Evidence
Large 4's safety story is genuinely more serious than most open-weight releases — and genuinely less documented than it should be.
On the serious side: Mistral ran a staged release. A moderated preview went to the public API, while cybersecurity leaders, vetted partners, and state authorities received the same model with reduced moderation and expanded cyber capabilities for real-world red-teaming. The company's case is that provider-level refusals can block legitimate vulnerability research and incident response — a real argument, and one that matters to the cyberdefense buyers Mistral is courting. Co-founder Guillaume Lample argues the lock-in point even more broadly: a customer holding the weights has a complete copy that no vendor can retire, while a closed model can disappear or be limited at any time. For cyberdefense, that argument carries weight.
On the less-documented side: Mistral has not published red-teaming findings or methodology, and no model card or system card existed as of October 7. The widely shared headline that the model "tried to escape its test environment" traces to a single sentence from Mistral's vice president of science, Pierre Stock, in a Reuters interview: the model "had tried to go beyond its testing environment, but that this was expected and the company was able to prevent it." The word "escape" is the headline's word, not Mistral's, and nothing published says what the model attempted, what stopped it, or how the episode was found. Treat the substance as unverified until Mistral documents it — and watch, after the weights go public, for fine-tunes that strip the safety tuning, and for how the reduced-moderation version actually gets used by its vetted holders.
There is also a policy backdrop worth knowing. The US administration temporarily restricted distribution of OpenAI and Anthropic models in June over cyberattack concerns, and the White House has reportedly asked American labs to withhold unreleased models even from the UK's AI Safety Institute. That is the world Mistral is selling into: governments worried about frontier cyber capability, and a European lab offering them a model they can hold, inspect, and run themselves. Whether that offer is reassuring or alarming depends on the licence, the red-teaming evidence, and the controls — which is why checks 2 and 4 belong together.
Check 5: The Practical Reality — Can You Actually Use It?
Strip away the launch narrative and ask the buyer's questions. Can you run it? At a trillion parameters, probably not on hardware you own — and possibly not on hardware your university owns. Can you afford the API? The pricing needs a second look given the launch-post versus model-page discrepancy. Does it beat what you already use? On coding, the vendor numbers say no — GLM-5.3 and Kimi K3 lead it, and closed models lead further. On cyber, geospatial, and agentic work, Mistral claims a top open-weight ranking, and the released weights will let independent testers confirm or deny it.
For regulated European buyers, the calculus is different, and it is the calculus Mistral is really targeting. The pitch is sovereignty: a frontier-scale model trained on European soil, in European datacenters, available as weights you control, from a company that counts Airbus, ASML, and HSBC among its enterprise customers. Europe's other open-weight bet, Aleph Alpha's Kolibri, took the same sparse-expert route at about a thirteenth of the size. The two releases together are the clearest test yet of whether European labs can field models that regulated buyers will actually choose — and Large 4's weights release is the moment that test starts producing evidence instead of press releases.
For everyone else, the practical advice is simple: test it yourself and trust your own numbers over any launch chart. Run it on your workloads, compare it against the model you already use, and read the licence before you build. All three become possible the day the weights land. Until then, the preview is the product — and previews are for evaluating, not for betting on.
How to Compare Large 4 Against What You Already Use
The most useful thing you can do while waiting for the weights is to fix your baseline. Pick the model you actually use today for coding work and check it against the best AI coding assistants landscape: if Large 4's 61.7% DeepSWE figure holds after independent testing, it still trails the leaders on that board, so a coding-first buyer should need a specific reason — sovereignty, self-hosting, or cyber capability — to switch. For general comparison discipline, the same rule from our ChatGPT vs Claude vs Gemini comparison applies here: same conditions, or no comparison. Vendor charts, preview documentation, and launch keynotes are not conditions. The weights, the licence, and independent replications are. Write down your current model's scores on your own workloads now, so that when Large 4's weights land, you are measuring against reality instead of against a launch chart.
What Happens Next
Two tracks are already arranged. The Mistral weights watch starts checking from October 27 at midnight PKT, every six hours, and stays silent until a real drop — official weights plus licence from Mistral or Hugging Face — at which point it reports once with the drop date, source links, licence terms, and download status, then stops. If nothing has landed by the end of October 31, it sends one closing note and stops. Note the date honesty built into that watch: October 27 is the Reuters-reported date, October 31 is the Hugging Face page estimate, and neither is Mistral-confirmed — which is exactly why this preview says "any day" instead of naming a day.
When the drop is real, the full Large 4 guide follows: the licence text quoted, the independent benchmarks checked against the vendor claims, the download and hardware reality tested, and the comparison against Gemini 4 Argon, Kimi K3.1 Pro, and the rest of the frontier updated with evidence instead of announcements. Until then, this preview is the record — and it will be corrected the day the weights land if any of it turns out wrong.
Frequently Asked Questions
When will Mistral Large 4's weights actually be released? Mistral says by the end of October. Reuters reported October 27; Mistral's Hugging Face page now estimates October 31. Neither date is Mistral-confirmed as fixed. Treat late October as the window, and the Hugging Face checkpoint as the proof.
Is Large 4 open source? Not in any confirmed sense yet. Large 3 shipped under Apache 2.0, but Large 4's weights are expected under a custom Mistral licence whose text has not been published. The docs badge reads only "Open." Read the licence text before building anything commercial on it.
How big is Large 4, really? About 1.05 trillion total parameters, with roughly 49–52 billion active per token — reports disagree on the active figure. It is too large for a desktop or laptop, and possibly impractical for some universities. The API preview is the only way to use it today.
Does Large 4 beat GPT and Claude? Not overall, on the numbers available. It trails on coding (61.7% DeepSWE v1.1 vs around 69% for GLM-5.3 and Kimi K3, around 74% for closed leaders) and claims strength in cyber (82% reproduce-and-patch, 93% Cybench) that still needs independent replication. Independent scores place it as a top Western open-weight model, not the best model in the world.
What is the reduced-moderation version? A version of Large 4 with fewer safety restrictions and expanded cyber capabilities, given to cybersecurity leaders, vetted partners, and state authorities for red-teaming before the public weights release. No red-teaming findings or methodology have been published yet.
What should I do on release day? Check five things in order: the Hugging Face checkpoint exists; the licence text and its commercial terms; independent benchmark replications; the safety documentation; and the real hardware and pricing requirements. This preview's five checks are the list.
Sources
- France's Mistral launches AI model it says outperforms some Chinese rivals
- Mistral Large 4 'Le Chonk': Europe's Trillion-Parameter Bid to Break the US-China AI Grip
- Mistral's "Le Chonk" Has 1 Trillion Parameters but Uses Only 49 Billion
- Mistral Large 4 Preview Pitches Open Weights Against Lock-In
- Mistral Large 4 Is Open Weights in Name and API First in Practice
- Mistral Large 4: a cyber model with an open-weights deadline
Reader reviews
Loading reviews…
