Mistral Large 4 Preview Scores 38 on Artificial Analysis; Open Weights Due Oct. 27

Bottom line: Artificial Analysis scores the Mistral Large 4 Preview at 38, tied with GPT-6 Luna. VentureBeat reports open weights are scheduled for October 27, 2026. Reports differ on how many GPUs were used for training.

Thirty-eight is a middle-of-the-pack number wearing a national headline. Unite.AI’s Oct. 6 account of Artificial Analysis says Mistral Large 4 Preview scores 38 on Intelligence Index version 4.3.2. The same writeup says that score is comparable to GPT-6 Luna (max) at 38 and one point behind DeepSeek V4.1 Flash (max) at 39, and that Artificial Analysis describes it as the most intelligent model from outside the United States and China. Rank on the firm’s comparison class: 64 of 225, against a class median of 26. Both sentences can be true. “Best outside two countries” and “64th in this table” are not the same claim. If you only keep one, keep the rank. The geographic framing is a press sentence. The rank is a position.

VentureBeat the same day says the weights are planned for Oct. 27, after about three weeks of testing, under a custom Mistral license. Unite.AI says “by the end of October,” and that the Artificial Analysis page still classifies the preview as proprietary because the weights are not public. TechCrunch says it is not an open-weight model yet. It is a public guardrail endpoint, with weights in three weeks after safety testing. Oct. 27 and “end of October” can be the same week. They are not the same sentence.

We already wrote a benchmark that a frontier model barely cleared in Opus 5.5 and the 34.8 percent on Argo-Bench. We wrote a price cut that was not a capability jump in Opus 5.5, Sol, and the halved rates. We wrote a model that shipped to security partners before the public call in Gemini 4 Argon. Large 4 is a fourth shape: a preview score, a closed weight file, and a date that still depends on which desk you ask.

What the Mistral Large 4 score of 38 is made of

Unite.AI, citing Artificial Analysis’s methodology for index version 4.3.2, says the score combines ten evaluations: AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity’s Last Exam, GDP.pdf, CritPt, AA-Omniscience, and AA-LCR v1.1. The weights are agents 30 percent, coding 20 percent, scientific reasoning 20 percent, general 30 percent. Artificial Analysis estimates a 95 percent confidence interval of less than ±1 percent for this score. DeepSeek V4.1 Flash (max) at 39 and Large 4 at 38 sit inside a band the firm itself treats as tight. Luna (max) at 38 is a tie, not a win.

The page details Unite.AI repeats are more useful than the country ranking. The preview is listed as a reasoning model, released Oct. 6, 2026, text and image in, text out, served through two API providers. On the index run it generated 200 million output tokens, against a class median of 81 million. Output speed, measured through Mistral’s API, was 116.1 tokens per second against a median of 86.1. Time to first token was 1.46 seconds against a median of 3.81 seconds. It is a talkative, reasonably fast preview. Talkative is not the same as cheap, which is the next section.

On GDP.pdf, a document-reasoning eval, Unite.AI says the preview scores 19 percent, on par with MiMo-V2.6-Pro at 19 percent. That is a single cell, not the index. They are different instruments. The index already includes GDP.pdf. Quoting the cell again as if it were a second verdict double-counts it.

Context, per the same analysis: 512,000 tokens. Parameter count, agreed by VentureBeat and Unite.AI: 1 trillion total, 49 billion active. Nickname Le Chonk. The training-cluster number is not.

4,000 GPUs, or 3,800

VentureBeat says Large 4 was trained from scratch over roughly two months on 4,000 Nvidia Grace Blackwell GPUs in Mistral’s own European data centers. Unite.AI’s recap of the launch post says 3,800 NVIDIA Grace Blackwell GPUs in the company’s own European datacenters. Two hundred GPUs is a rounding error in a press metaphor and a real difference in a capacity plan. Until Mistral’s own post is the sentence you quote, say the outlets disagree and cite both. Unite.AI links it.

What does agree: more than 160 languages, including every official language of the European Union. VentureBeat’s focus list, attributed to the company, is coding, cybersecurity, finance, manufacturing, and visual grounding. CNBC says cyber, coding, manufacturing, finance, and multimodal. Close enough to treat as one pitch with two edits, not as two products.

Mistral says that once the core parameters are released, Large 4 will rank among the top open-weight models on aggregate benchmarks, and that it is the strongest open-weight model developed outside China by a “substantial margin.” The same paragraph says the model still lags the frontier in coding, and that capabilities will improve as training capacity scales after the Series D. A 38 that lags the frontier in the category weighted at 20 percent of the index is a coherent result. It is not a scandal. It is a reason not to buy the model for a coding migration on the strength of the country ranking.

The Series D figures need the same care as the GPU count. CNBC: a 3 billion euro round, $3.4 billion, in September, at a 21 billion euro valuation. TechCrunch: Series D last month at a 21 billion euro valuation, about $24.39 billion. VentureBeat, citing Reuters: valuation roughly $24 billion. The 3 billion euros is the round. The 21 billion euros is the valuation. The dollar conversions of the valuation do not match each other to the cent, and they are not the round size. If a chart puts “$3.4 billion” and “$24 billion” in one cell, it has mixed a raise with a valuation.

The weights are the product. The endpoint is the preview.

TechCrunch is the bluntest on access. Large 4 is not open-weight yet. For now it is a public guardrail endpoint. Weights in about three weeks, after safety testing. The piece reports that Mistral will work with trusted partners and governments so the open-source weights can be used to defend, not to attack. The condition that is on the page is a defend-not-attack constraint during the wait, plus a partner-and-government testing period. That is a release process. It is not a weight file you can hash today.

VentureBeat dates the publication to Oct. 27 and says the license is expected to be a custom Mistral license. It also says the company did not provide API pricing in the materials VentureBeat reviewed. Unite.AI, via the Artificial Analysis page, does have prices, and they are discount prices: a launch discount for the first two weeks cuts token prices to $0.68 and $2.09 and brings cost per task to $0.57. That is still above GLM-5.3-Flash at $0.25 per task and DeepSeek V4.1 Flash (max) at $0.27. Artificial Analysis characterizes the preview as more than four times the cost per task of similar-intelligence open-weights models. If you are comparing invoices, the two-week discount is the number with a source, and it already loses to the open-weight models sitting one point away on the index.

So the economic reading of a 38 is uncomfortable and specific. You can buy a score that ties Luna, from a European lab, at a task cost more than double the Flash-class models at 39 and below, and you cannot yet run the weights yourself. The sovereign-infrastructure pitch in VentureBeat, including zero-data-retention options, is the reason a buyer might pay that anyway. Guillaume Lample told VentureBeat that customers need deployment, infrastructure, customization, and engineering support, not just model access. The company says it supports more than 125 global enterprises, including Airbus, ASML, and HSBC. That is a services business sitting on top of a weight release. It is a coherent business. It is not what the Intelligence Index measures.

TechCrunch’s interview, with Stock, happened while benchmark results were still described as pending. Mistral hoped to be best in class among open-weight models, especially outside China, and maybe to beat closed models in specific customer areas where multimodal input helps. Hope, in an interview published the same day as a 38, is not a second score. The 38 arrived. The “best outside China” line is Mistral’s claim about the open-weight field after the weights exist. The weights do not exist yet. CNBC’s “substantial margin” is the same tense: once the parameters are released.

DeepSWE is a chart, and then a different chart

VentureBeat is where the coding caveat gets numbers. Mistral’s materials, as that piece reads them, show Large 4 at 62 percent on DeepSWE v1.1. The comparison chart lists Reflection AI’s Beam at 44 percent, Qwen 3.8 Max at 51 percent, DeepSeek V4 Pro 0813 at 57 percent, and GLM-5.3 at 61 percent. Configuration is the whole argument. Reflection’s own Beam announcement, VentureBeat notes, reports 44.4 percent for Beam, 51 percent for Qwen 3.8 Max, and 61 percent for GLM-5.3 under its setup. Artificial Analysis separately reports 57 percent on DeepSWE for DeepSeek V4 Pro 0813 when paired with the Codex agent harness. Those figures are traceable. They are not one leaderboard.

The live DeepSWE board, selecting the best published configuration for each model, is a different picture. VentureBeat says it currently puts GLM-5.3 at about 69 percent and Kimi K3 at about 69 percent, with GPT-6 Astra, Gemini 3.8 Flash, and Claude Opus 5 around 74 percent. A 62 percent preview result looks competitive against the chart Mistral printed. It does not look competitive against the best published configs on the live board. If your migration is long-horizon software engineering, quote the live board, not the launch chart. If your migration is “we need a European weight file in late October for a regulated workload,” the DeepSWE gap is a risk item, not a reason to pretend the launch chart was the only measurement.

62 percent, Mistral’s chart, DeepSWE v1.1. About 74 percent, the closed models on the live best-config board, as VentureBeat summarized it on Oct. 6. The distance between those two is the coding lag CNBC stated in words.

What not to conclude

Do not treat “the most intelligent model outside the US and China” as a durable fact. That is Artificial Analysis’s framing of one index version, on a preview, on Oct. 6. Rank 64 of 225 is the more stable way to say the same measurement. A tie with Luna at 38 and a one-point gap to DeepSeek V4.1 Flash (max) is the comparison that survives a redraw.

TechCrunch, VentureBeat, and the Artificial Analysis classification agree that they are not, today. Oct. 27 is VentureBeat’s date. End of October is Unite.AI’s. A custom license is not Apache-2.0. A defend-not-attack testing period with governments is a constraint, not a mirror of DeepSeek’s release shape.

The model is not simply “expensive” in the abstract. It is expensive relative to open-weight models of similar index score, on Artificial Analysis’s cost-per-task, during a discount window, at $0.57 against $0.25 and $0.27. After the weights ship, the cost model changes again, because you can run them. Those are two future invoices. This article only has the preview invoice.

The useful analysis is short. Large 4 is a 1-trillion-parameter preview, 49 billion active, 512k context, scored 38 on a ten-test index, tied with a GPT-6 Luna max setting, one point behind a DeepSeek Flash max setting, and more than four times the task cost of similar open weights while its own weights sit until late October. The GPU count in the secondary coverage does not match. The coding chart the company printed does not match the live board VentureBeat checked. None of that makes the preview fake. It makes the headline “Europe’s most intelligent model” a smaller statement than it sounds, and it makes Oct. 27 the date worth putting on a calendar instead of the score.

What this means for you

If you are considering self-hosting, wait for the official weights release and model card, then check license terms and hardware requirements. For API use, test the preview endpoint against your own tasks; a single aggregate score does not tell you how it handles your workload.