Mistral Large 4 Is Europe's Best Model. It Is Not the Best Open Model, and It Is Not Open Yet.

Mistral launched a public preview of Mistral Large 4 this morning, October 6, 2026, and gave it a nickname before it gave it a license. "Unofficially ML4, very officially: le Chonk," the blog says. It is a 1 trillion-parameter mixture-of-experts model, Mistral's largest, trained from scratch in the company's own European datacenters, and its first new model since May. The preview API is live on Mistral's platform at $1.36 in / $4.18 out per 1M tokens, half price for the first two weeks.

Artificial Analysis scored it the same day: 38 on the current v4.3.2 Intelligence Index, which AA calls "the most intelligent model from outside the US and China." That sentence is true, and the qualifier is doing all the work. The strongest open-weight models are Chinese, and at least five of them sit above 38 on AA's own leaderboard, Kimi K3 among them. Mistral's own claim is narrower than the headlines: "the best open weights model from US or Europe on aggregated benchmarks."

And "open weights" is, as of today, a promise. There are no weights, no named license, no knowledge cutoff, no speed data, and the blog says the checkpoint "continues to improve rapidly as we refine it," so every number below describes the preview as it stood this morning. We read the blog, the model card, the docs changelog, the pricing page, the reasoning docs, Artificial Analysis, and the launch-day press, some of which got the basics wrong. Here is what held up.

A Trillion Parameters, Sold Like a Closed Model

The active-parameter count is Mistral's, and it is per token. The blog and the 13:02 UTC launch tweet both say 49 billion active parameters out of 1 trillion; the model card rounds the total to 1.05T. Update (Oct 16): Mistral's blog and model card now say 52 billion, and its Hugging Face placeholder defines the pair as 49 billion per token and 52 billion including embeddings and output layers; the full note is in the update below.

The license column says "Open." The license field is empty. Mistral's models overview prints "Open" next to Large 4 where Large 3 says "Apache 2.0" and Medium 3.5 says "Modified MIT." No Mistral page names the license. VentureBeat and The New Stack write "custom Mistral license"; that is their inference, not a Mistral statement, so we print "not yet published." The precedent cuts both ways. Mistral Large 3 shipped Apache 2.0 last December, and Pierre Stock, Mistral's VP of Science, told TechCrunch the company will "work with trusted partners and governments to make sure that the open source weights can be used to defend, but not to [perform] malicious attacks," which is not a sentence anyone writes about Apache 2.0.

The weights date is "end of October," full stop. The blog says "Weights drop end of this month" and, lower down, "We will release the weights by the end of the month, along with more details on the architecture, additional benchmarks, and our post-training methodology." The tweet says "Open weights release end of October." VentureBeat, TNW, DW and The New Stack all print October 27; TechCrunch says "in just three weeks." Neither date appears in anything Mistral has published.

Context depends on whom you ask. Mistral's card and changelog say 1M tokens. Artificial Analysis's launch article says 512k and its model page 524k, and Mistral offers no explanation. Update (Oct 16): the gap is Mistral's EU endpoint, which carries half the context at a 10% premium; details in the update below.

Reasoning is a switch, not a dial. The reasoning docs list reasoning_effort values of high ("The response includes a full thinking chunk before the final answer, at the cost of increased token usage") and none. There is no low, medium or max, and no documented default. Simon Willison confirmed the same two levels against the API this morning. AA's 38 is for the reasoning configuration.

3,800 GPUs, not 4,000, and Mistral never writes a part number. The blog: "ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own datacenters in Europe." Stock told TechCrunch 4,000, "which is two to three times less than our Chinese competitors, and significantly less than the closed source competitors," and VentureBeat, TNW and The New Stack printed 4,000 as fact. The Register printed 3,800 and worked it out to about 52 NVL72 racks. The "two months" and "10 megawatts" figures are TNW's, not Mistral's.

The Benchmarks, Sorted by Who Reported Them

House rules: independent numbers first, stated as fact; Mistral's numbers labeled as Mistral's. Artificial Analysis has not re-versioned its index since v4.3.2 on September 19, so these scores compare directly with the ones in our September posts.

Thirty-eight is a generational jump for Mistral and a mid-table score for everyone else. Mistral Large 3, released December 2, 2025, scores 9 on the same index. Willison calls 38 "a huge improvement on last December's Mistral Large 3, which scored 9 on AA" and, in the same post, "not a Fable-class model." AA's own comparison set is DeepSeek V4.1 Flash and GPT-6 Luna. The top of the same index this morning is Claude Opus 5.5 (max) at 58 and Claude Sonnet 5.5 (max) at 56.

"Most intelligent model from outside the US and China" is AA's exact wording, and the qualifier matters. On AA's leaderboard, same v4.3.2 index, at least five Chinese open-weight models score higher: MiMo-V2.6-Pro 46, GLM-5.3 (max) 45, Kimi K3 (max) 44, GLM-5.3-Flash 42, and DeepSeek V4.1 Flash (max) 39. Mistral's blog says ML4 is "significantly outperforming any open-weight model developed in the US or Europe," and on AA's chart that holds: the best US open-weight entry, Thinking Machines Lab's Inkling (xhigh), scores 25. "Strongest open model," which some headlines implied, is false on the index AA publishes, and Mistral itself never says it.

Cost per task is where the open-weight comparison hurts. ML4 generated 200M output tokens to run AA's index, against a median of 87M, so its $1.13 per task is, in AA's words, "Over 4x the Cost per Task of similar-intelligence open weights models": GLM-5.3-Flash at 42 costs $0.25 per task and DeepSeek V4.1 Flash (max) at 39 costs $0.27. The discounted $0.57 does not close that gap. Closed models make it worse. GPT-6 Luna (max), the other 38 on the index, costs $0.07 per task, and GPT-6.1 Sol (max) scores 52 for $0.72.

The cyber claim is accurate and selective. Mistral's blog says ML4 "ranks among the top five models globally and leads open-weight models developed outside China by a wide margin" on AA's Cyber Index. The chart today: Grok 4.7 (xhigh) 56, MiMo-V2.6-Pro 56, GPT-6 Luna (max) 53, GLM-5.3-Flash 50, Mistral Large 4 Preview 50. Fifth by position, tied fourth by score. The 82% that Mistral and The Decoder lead with is one of the index's three equally weighted tests; the blog does not print the other two. The context Mistral supplies is fair: "Several leading closed models, including Claude Opus 5.5 and GPT-6 Astra, score near zero" on the vulnerability-reproduction test because they refuse it. AA's projection: "Once its weights are released, it will rank among the top three open weights models on the Cyber Index."

Mistral also quotes one AA figure of its own: 1,393 Elo on AA-Briefcase, "ahead of DeepSeek V4 Pro." AA's board shows the same number.

Now the vendor table. These are Mistral's self-reported numbers from the launch blog and the materials it gave VentureBeat, TNW and The New Stack. Harnesses and competitor effort levels are not stated; VentureBeat notes that "Mistral supplied preliminary results."

Four honest observations Mistral's marketing does not lead with:

• 49.8% is arithmetic, not a leaderboard entry. (61.7 + 59.4 + 28.3) / 3 = 49.8. Artificial Analysis publishes a Coding Agent Index v1.5 built from exactly those three benchmarks, and ML4 is not on it; VentureBeat confirmed this morning that ML4 "does not yet appear in Artificial Analysis' public evaluations or the DeepSWE leaderboard." Print it as Mistral's mean or do not print it.
• 28.3% on Terminal-Bench 4 is the row the sentence hides. AA's own runs put Claude Sonnet 5.5 at 64%, Claude Opus 5.5 at 59.6% and Gemini 4 Argon at 57% on Terminal-Bench 4.0. Harnesses differ, so treat it as a gap rather than a precise one. It is a large gap.
• The DeepSWE comparison set is curated. Mistral's 61.7% beats the open models it chose. VentureBeat's read of the live DeepSWE board has GLM-5.3 and Kimi K3 around 69%, and GPT-6 Astra, Gemini 3.8 Flash and Claude Opus 5 around 74%.
• The Harvey 15% is not Harvey's number. The press attributes it to Mistral's materials; the benchmark is run by Vals AI, which had not published its own figure this morning. See the update.

The Half-Price Fortnight Has No End Date

Mistral's changelog says "Launch pricing: 50% off for 2 weeks." The model card and pricing page show the sale prices, $0.68 in / $0.07 cached / $2.09 out, next to the list prices. No Mistral page prints an end date; two weeks from October 6 is on or about October 20. Budget against $1.36 / $0.14 / $4.18 and read the card again on the 20th.

Three press errors, all on day one. VentureBeat wrote that "the company did not provide API pricing in the materials reviewed by VentureBeat," while the blog and the model card both printed prices this morning and Artificial Analysis quoted them. The Decoder printed $0.68 / $2.09 as the price without saying it is the discount. And the 4,000-GPU figure that VentureBeat, TNW and The New Stack print as Mistral's is a round number from an interview; the primary source says 3,800.

Against Mistral's own lineup. Mistral Large 3 (mistral-large-2512, December 2025, 675B total / 41B active, Apache 2.0) stays on sale at $0.50 / $0.05 / $1.50 with no deprecation notice, so Large 4 lists at 2.7x the input and 2.8x the output price of its predecessor. It undercuts Mistral Medium 3.5 on output ($1.50 / $7.50). Cached input is 10% of input across all three, with no cache-write charge printed.

What is not on the price list. No Batch or Priority tier price, no EUR figure (the card has a currency toggle; the pricing page prints USD only), no image-token price, no long-context surcharge. The card lists /v1/batch as a supported endpoint with no batch price beside it. Dashes for all four until Mistral prints them.

Where You Can Run It

• Mistral API and Studio: yes, public preview, through /v1/chat/completions, /v1/conversations and /v1/batch, with function calling, structured outputs and document QnA. The tweet: "Available to all via API today. Working with cybersecurity partners privately."
• OpenRouter: yes, listed at 13:42 UTC, under two hours after the blog went up.
• AWS Bedrock, Azure AI Foundry, Google Vertex AI: - (none lists it, and Mistral's blog names no cloud partner).
• Le Chat and Vibe: - (no source says whether ML4 is in the consumer app, or on which plan).
• Self-hosting: future tense in the blog, "it will be able to run on private cloud or on-premise," and it depends on the weights.

The preview "is served on that same infrastructure," meaning Mistral's own European datacenters, with what VentureBeat describes as zero-data-retention and sovereign options. Vetted cybersecurity partners and state authorities get "the same model with reduced moderation and expanded cyber capabilities." There is no system card, no bio or chem evaluation and no third-party safety audit. The one safety anecdote on record is Stock telling Reuters, as relayed by The New Stack, that during testing the model "tried to go beyond its testing environment," that this was expected, and that software controls contained it.

Verified vs Unconfirmed: The Scorecard

Who Should Use It

Use Mistral Large 4 if your constraint is jurisdiction rather than leaderboard position: a European deployment Mistral says it "operates end-to-end, independently of other digital service providers and under European law," zero-data-retention endpoints, more than 160 languages including every official EU language, and cyber-defense work where Claude Opus 5.5 and GPT-6 Astra refuse the task. At $0.57 per AA index task during the discount it is a cheap way to find out whether the preview fits, provided you re-read the card around October 20.

Skip it for cost-sensitive production: at list it is over 4x the per-task cost of GLM-5.3-Flash and DeepSeek V4.1 Flash, and 16x GPT-6 Luna at the same index score. Skip it for coding agents (28.3% on Terminal-Bench 4 by Mistral's own count, and no AA coding entry). And skip it for anything you cannot re-benchmark in three weeks, because the checkpoint "continues to improve rapidly," the weights and license do not exist yet, and there is no knowledge cutoff, no speed figure and no system card to put in a procurement file. If you need weights you can download today, that is Kimi K3 or Gemma 4, not le Chonk.

The frontier is elsewhere: Claude Opus 5.5 (max) leads the v4.3.2 index at 58, Claude Sonnet 5.5 (max) is at 56, and Gemini 4 Argon (high), GPT-6 Astra (max) and Claude Fable 5.1 (max) are at 53. Mistral Large 4 at 38 is the best model Europe has produced, and Willison's estimate of about six months behind the frontier reads right. Our model comparison has the current standings side by side.

Update, October 18, 2026

• Oct 17 and 18: DeepSeek moved deepseek-flash to a new build, DeepSeek-V4.1-Flash-1017, under the same name, and Artificial Analysis rescored it at 45 on the same v4.3.2 index, up from the 39 this post lists as ML4's nearest neighbor. The gap between Mistral Large 4 and DeepSeek's Flash is now seven points, not one, and the September build it was compared with is no longer served on DeepSeek's API. See our DeepSeek V4.1 Flash 1017 post.
• Oct 8: Artificial Analysis added trusted-access models to its Cyber Index. GPT-6 Sol (Daybreak Blue, max), available only through OpenAI's Daybreak program, scores 69 and now leads; Mistral Large 4 Preview is sixth of 20 by position and still fifth among publicly available models. Mistral's "top five" was accurate when written.
• Oct 9: Tom's Hardware, reading AA's per-test chart, puts ML4 at 51% on CWE-Bench-AA and 16% on DeepsecBench-AA beside the 82% on CyberGym-E2E-AA, and names the same five Chinese open-weight models above it on the index.
• Oct 15: AA scored Meta's Muse Spark Open (xhigh), a US open-weight model whose weights are on Hugging Face, at 39 on the same index, one point above ML4. Mistral's "best open weights model from US or Europe" line now has a US challenger that is actually open.
• Oct 16: The 50% discount is still live on all three of Mistral's OpenRouter endpoints, ten days after launch, with no end date printed anywhere. Those endpoints also settle the context dispute: Standard and ZDR carry 1,048,576 tokens, the EU endpoint 524,288 at a 10% premium (about $1.50 / $4.60 list), which is the figure AA (524k) and Vals AI (512k) print without saying which endpoint they benchmarked. Mistral's blog now reads "52 billion active parameters" where the Wayback Machine's capture at 23:01 UTC on October 6 reads 49 billion; the model card says 52B, and Mistral's Hugging Face placeholder, mistralai/Mistral-Large-4-1T-A52B, defines the pair as "49 billion active parameters per token (52 billion including embeddings and output layers)." That placeholder carries a countdown to an October 31 ETA, 839 people waiting, no files and no license field; it is the only weights date Mistral has published. No weights, no license, no GA announcement, and AA still labels the preview "Proprietary model" on an index that is still v4.3.2. Independent scores that did not exist on launch day: Vals AI's Vals Index 48.11% (#35 of 48), Harvey's Legal Agent Benchmark 15.83% (#9 of 79, the benchmark behind the press's 15%), Vibe Code Bench v1.1 78.40% (#28 of 112); on AA, AA-Briefcase v1.1 1,393 Elo (#56), GDPval-AA v2.1 1,422 Elo (#67), Harvey LAB-AA v1.1 rank 12 at 3.1%, and AA-Omniscience index -5 with a 42% hallucination rate. ML4 is still not on AA's Coding Agent Index v1.5.

Sources

• Mistral: Introducing Mistral Large 4
• Wayback Machine capture of the Mistral blog, Oct 6, 2026, 23:01 UTC (says 49B active)
• Mistral model card: Mistral Large 4
• Mistral docs changelog (Oct 6 entry: public preview, 1M context, 50% off for 2 weeks)
• Mistral reasoning docs (reasoning_effort: high or none)
• Hugging Face placeholder: mistralai/Mistral-Large-4-1T-A52B (ETA October 31)
• Artificial Analysis: Mistral Large 4 makes France home to the most intelligent model outside the US and China
• Artificial Analysis: Mistral Large 4 Preview model page
• TechCrunch: Mistral's new 1T model aims to leapfrog closed and open rivals

*Not sure whether a sovereign-but-closed preview fits how you work, or whether you need open weights that exist today? Take the free 2-minute quiz and get matched. Then read the other API news from the same morning: OpenAI's Decisions API.*