Mistral Large 4 "Le Chonk": 1T Model, $1.36 Price, Weights Due Oct 27

Mistral Large 4 has been the most-watched open-weight release of the month for three weeks, and today is the day the weights are supposed to land. The model, which Mistral itself calls "very officially: le Chonk," has been available only through a paid API since its October 6 preview. The download that makes it genuinely open — the trillion parameters, the license, the architecture details — was promised for the end of October, and multiple outlets, citing Reuters, have pointed to October 27 as the date. As of this morning, the weights have not appeared on Hugging Face, and Mistral has not confirmed a time. That gap between the reported date and the company's own wording is the first of several worth keeping in mind, because this is a model whose launch post and independent tests tell noticeably different stories.

The short version: Large 4 is a 1-trillion-parameter mixture-of-experts model with about 52 billion active parameters per token, built for coding, cybersecurity, finance, legal work, and reading technical images. Its strongest results are independently verified — it tops one Artificial Analysis cybersecurity test and leads open-weight models on a legal benchmark — but its broad aggregate ranking is 32nd of 44 models, and it trails Kimi K3 on coding in Mistral's own chart. The API costs $1.36 per million input tokens and $4.18 per million output tokens. The weights are due today by one widely reported date, and by the end of the month by Mistral's own words.

What Mistral Large 4 actually is

Large 4, or ML4, is Mistral's largest and most capable model to date, in the company's own description. It is a natively multimodal mixture-of-experts system: text and images go in, text comes out. The headline numbers are about 1 trillion total parameters with roughly 52 billion active for any given token, according to Mistral's announcement and its documentation, which now agree on those figures. Earlier coverage cited 49 billion active parameters, and neither Mistral nor its docs has explained the change — a small discrepancy, but the kind that matters when you are sizing hardware for a self-hosted deployment.

The model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own datacenters in Europe. The training data spans more than 160 languages, including every official language of the European Union, which is central to the pitch: a European model, trained on European infrastructure, offered as an alternative to American closed models and Chinese open ones.

One detail changes how you should read every benchmark in this article. The reinforcement learning run behind Large 4 is still going. Mistral says the run currently uses about 3,000 GPUs, generates roughly 33 billion tokens a day, and is "showing no signs of saturation." The model you can call through the API today is not the model whose weights ship — every score below is a snapshot of a moving target.

What it costs — and the fine print

The standard API price is $1.36 per million input tokens and $4.18 per million output tokens. That is the rate that has applied since October 20, when a two-week launch promotion ended. During the promotion, from the October 6 launch, the price was half that: $0.68 per million input tokens and $2.09 per million output tokens, according to launch coverage that tracked the rate change. Anyone who benchmarked costs during the promotion and built a budget on those numbers is now underestimating the bill by a factor of two.

Cached input — tokens you have already sent and Mistral stores for reuse — costs $0.14 per million, about a tenth of the standard input rate. That is the number that matters for agent workloads that repeat the same context. Batch API pricing is reported at roughly half the standard rates, but Mistral's public pricing page does not spell out a batch figure for Large 4, so treat that as reported, not confirmed.

Two comparisons put the price in context. Against Kimi K3.1 Pro, which we covered on October 24 at $2 per million input and $10 per million output, Large 4 is cheaper on both lines. Against DeepSeek V4.1 Flash, at $0.15 per million input and $0.60 per million output off-peak (double at peak hours), Large 4 costs roughly nine times as much per input token at off-peak rates. That is not a like-for-like comparison — Flash is a smaller, faster model built for a different job — but it shows the range of the current open-weight market: Large 4 sits in the middle, neither the budget option nor the premium one.

There is no published price for self-hosting, because there is no hosted product to price yet. Once the weights ship, the cost becomes hardware: a trillion parameters have to sit in memory somewhere, and Mistral has not published hardware requirements. Most users will keep paying the API rate.

The benchmarks Mistral is showing

Mistral's launch post is built around a simple claim: Large 4 "already achieves performance competitive with the strongest open-source models globally, while significantly outperforming any open-weight model developed in the US or Europe." The numbers underneath that claim are a mix of company-run tests and privately run independent ones, and they need to be separated.

On coding, Mistral reports 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4, for a combined Coding Agent Index of 49.8% — ahead, it says, of DeepSeek V4 Pro 0813 and Qwen3.8 Max. But look at Mistral's own DeepSWE chart and the tallest bar is not Large 4. It is Kimi K3, at 69%, seven points ahead. Large 4 edges GLM-5.3 by a single point, which is well within the noise of harness choice. On the live Datacurve DeepSWE leaderboard, which runs every model in the same setup, GLM-5.3 and Kimi K3 both reach 69% and DeepSeek V4 Pro reaches 63%. Large 4 is not listed there yet. And the DeepSWE, Terminal-Bench, and SWE-Atlas scores were, by a footnote on Mistral's own charts, evaluated privately by Artificial Analysis ahead of that harness's public launch — so nobody outside Artificial Analysis and Mistral can check them until the harness is public.

On agentic workflows, Mistral reports 59.9% on AutomationBench, a set of 657 business workflows across Gmail, Google Sheets, Slack, and Salesforce, ahead of Kimi K3, MiMo-V2.6-Pro, and DeepSeek V4 Pro. On visual grounding — locating a specific object in an image from a text description — it reports 42% on Dense 200, one point ahead of GPT-6 Astra at 41%.

The human evaluations are more interesting than the automated ones. Mistral commissioned a blind evaluation from Surge AI in which professional annotators rated coding outputs from 1 to 5, with model identities hidden. Large 4 placed second of five at 3.74, ahead of GLM-5.3 (3.60), Kimi K3 (3.59), and GLM-5.2 (3.40), and behind only Claude Opus 5 (4.22). Kimi K3 beats Large 4 by seven points on DeepSWE and loses to it on human preference. Passing tests and writing code people like reading are not the same skill.

Cybersecurity is the real story

If there is one vertical where Large 4 earns its launch-post confidence, it is cybersecurity — and unusually for this launch, the evidence is independently verified. On Artificial Analysis's Cyber Index, an independent evaluation of how well models find and fix security flaws in real software, Large 4 scores 50%, placing it in the top five of the 18 models tested. Only Grok 4.7, MiMo-V2.6-Pro, and GPT-6 Luna score higher, and GLM-5.3 Flash ties it. Mistral's claim that it "leads open-weight models developed outside China by a wide margin" is consistent with that chart.

The standout is one component of the index, CyberGym-E2E, which asks a model to reproduce a real vulnerability in open-source software and then patch it. Large 4 scores 82% there, first of the 18 models tested, ahead of MiMo-V2.6-Pro at 79% and GPT-6 Luna at 78%. Mistral also reports 93% on Cybench, a set of 40 security competition challenges, though nobody has reproduced that score. On Lakera's public B3 AI Security Benchmark, which measures resistance to prompt-injection attacks, Mistral reports 93.3%, and says it sees no higher scores among competitors.

The refusal data is what makes this more than a benchmark story. On CyberGym-E2E, Artificial Analysis's chart shows Claude Opus 5.5 blocked on safety grounds for roughly 96% of tasks, and GPT-6 Astra blocked on 100%. Those models score near zero because the task gets refused, so their scores say nothing about what they could actually do. Mistral's argument is that defensive security work often starts with reproducing a flaw to prove it exists, and a model that refuses most of that workload — or whose provider can change its moderation rules mid-incident — is a liability. A security team running Large 4 on its own infrastructure controls the model's behavior, and the provider cannot take the capability away.

That argument cuts both ways, and Mistral knows it. Once the weights are public, attackers get the same capability. Artificial Analysis's index is deliberately defensive — models work from source code and are never asked to build a working exploit — so an 82% reproduction score is not 82% offensive capability. That is why Mistral is spending the three weeks before release red-teaming the model with cybersecurity leaders, vetted partners, and state authorities, who get a version with reduced moderation and expanded cyber capabilities.

What the independent tests say about everything else

Outside cybersecurity, the independent picture is mixed, and in places it contradicts the launch post. On finance, Large 4 scores 54.68% on vals.ai's Finance Agent v2, ranking 22nd of 75 models overall and 7th of 32 among open-weight models — solidly mid-pack, as DataCamp's independent analysis of the launch reports. On legal work, the picture is stronger: 15.83% on Harvey's Legal Agent benchmark, ranking 6th of 75 overall and first among 31 open-weight models. Sixth in the world at autonomous legal work still means completing about one task in six — no model should be taken as able to do legal work unsupervised.

On terminal-heavy coding, the independent number is worse than Mistral's. Vals.ai's independent run of Terminal-Bench 4.0 gave Large 4 22.73%, against the 28.3% Mistral reports, placing it 20th of 44. That is not evidence of anyone inflating anything — one number from one setup is not a ranking — but it is a reminder that a 5.6-point gap between a vendor's number and an independent rerun is normal, and that the vendor's number is the one that makes the launch post.

Then there is the aggregate. On vals.ai's broad Vals Index, Large 4 lands 32nd of 44 models overall and 9th of 16 among open-weight models, at 48.05%. It is hard to square that with "competitive with the strongest open-source models globally," at least on aggregate. The vertical results are real and they are not wrong; they just measure something narrower.

What "open weights" actually means here

Open-weight means the trained model parameters are publicly available. It does not mean the training data, training code, or anything else is released under an open-source license, and in Large 4's case it does not yet mean anything at all, because the weights are not out. Mistral's documentation describes Large 4 as open-weight but has not published the license terms. Until it does, commercial use, fine-tuning rights, and redistribution are all open questions.

The practical case for open weights is nonetheless real. A bank running Large 4 on its own servers never sends customer data to Mistral. A government agency controls the whole deployment environment. And if a provider goes down or retires a model version, self-hosted deployments keep running. Mistral says Large 4 uses the same training, customization, and reinforcement learning environment it offers enterprise customers through Mistral Forge, so organizations that want to fine-tune the model on their own data have an obvious route — once the weights and the license exist.

The context window is another unresolved number: 1 million tokens according to Mistral's documentation, 512,000 according to vals.ai. Check the docs on the day you build, not the day you read this.

How Large 4 compares

The table makes the shape of the race clear. On coding, the Chinese open-weight models are ahead on the one leaderboard that runs everyone in the same setup. On cybersecurity and legal work, Large 4 is the strongest open-weight option with independent evidence behind it. On general capability, it sits in the middle of the pack.

One related note for readers following the open-weight race: Kimi's newer model, Kimi K3.1 Pro, launched on October 24 with a $2 per million input and $10 per million output API price and open weights promised by November 4. We covered it in detail in our Kimi K3.1 Pro article. Mistral's other recent release, the Voxtral 3 voice model, is a different product entirely — a phone-ready voice agent at $0.04 a minute, covered in our Voxtral 3 review.

Pros and cons

Pros: the strongest independently verified cybersecurity results of any open-weight model, including a first-place 82% on CyberGym-E2E; first among open-weight models on Harvey's Legal Agent benchmark; a European deployment option under European law; a 1-million-token context window if the documentation figure holds; and human-evaluated coding quality that beats every open-weight rival in the blind Surge AI evaluation.

Cons: the weights, the license, the full architecture, and the hardware requirements are all still unpublished; the broad aggregate ranking is 32nd of 44; coding trails Kimi K3 on DeepSWE by seven points; the independent Terminal-Bench rerun came in 5.6 points below Mistral's number; the API price is roughly nine times DeepSeek V4.1 Flash's off-peak input rate; and the reinforcement learning run is still going, so the model you test today is not the model whose weights ship.

What we still don't know

Final benchmark performance: the reinforcement learning run is still in flight, so every score above may change. Independent coding, vision, and knowledge-work results: beyond the Artificial Analysis Cyber Index and vals.ai, nothing has been reproduced. Pricing: the launch announcement and the documentation have disagreed on rates, and the preview rates may not carry forward. License terms: you cannot build a product on "open-weight" without a license. Hardware requirements: not published. Full architecture: promised with the weights. Context window: 1 million tokens per Mistral's docs, 512,000 per vals.ai. October 27 is the date to watch — that is when the weights are supposed to ship and independent researchers can test the model Mistral actually releases instead of a preview that is still training.

FAQ

When do Mistral Large 4's weights come out? Mistral says "by the end of the month." Multiple outlets, citing Reuters, have reported October 27, 2026. As of the morning of October 27, no weights repository had appeared on Hugging Face and Mistral had not confirmed a time.

Is Mistral Large 4 really open source? No — it is open-weight, which is not the same thing. The trained parameters are promised for public release, but the training data and code are not, and the license terms have not been published. Until they are, commercial use, fine-tuning rights, and redistribution are open questions.

How much does the Mistral Large 4 API cost? The standard rate is $1.36 per million input tokens and $4.18 per million output tokens, with cached input at $0.14 per million. A two-week launch promotion at half those rates ran from October 6 and ended on October 20. Batch pricing is reported at roughly half the standard rates but is not confirmed on Mistral's public pricing page.

Can I run Mistral Large 4 on my own hardware? Not yet — the weights are not out. Once they are, expect a multi-GPU cluster: a trillion parameters have to sit in memory somewhere, and Mistral has not published hardware requirements. Most teams will keep using the API.

Is Mistral Large 4 better than Kimi K3 or DeepSeek V4 Pro? It depends on the job. On cybersecurity and legal work, yes, with independent evidence. On coding, no — Kimi K3 leads on DeepSWE by seven points and DeepSeek V4 Pro leads on the live leaderboard. On general capability, it ranks 32nd of 44 on vals.ai's broad index. Pick the vertical, then pick the model.

Key takeaways

• Mistral Large 4 is a 1-trillion-parameter model with about 52 billion active parameters per token, in public preview since October 6, with weights reported for October 27 and promised by Mistral for the end of the month.
• The API costs $1.36 per million input tokens and $4.18 per million output tokens; the half-price launch promotion ended October 20, and cached input runs $0.14 per million.
• Independent tests back the cybersecurity and legal claims — first place on CyberGym-E2E, first among open-weight models on Harvey's Legal Agent benchmark — but not the coding headline, where Kimi K3 leads on DeepSWE and vals.ai's Terminal-Bench rerun came in below Mistral's number.
• On a broad aggregate, Large 4 ranks 32nd of 44 models, which does not match the launch post's claim of competing with the strongest open-source models globally.
• Do not build on it until the license, the context window, and the final checkpoint are published — all three are still unpublished or disputed, and the reinforcement learning run behind the preview is still going.

Large 4 is a real release with real numbers, wrapped in a nickname that started as a joke. The honest story is about choice: Europe now has a trillion-parameter open-weight model, trained on its own soil, aimed at companies that want control over their data. Whether it earns a spot in your toolkit depends on three things — the independent tests, the license, and your own results. Until the weights land, test the API yourself, and trust your own numbers over any launch chart.