Mistral's Large 4 ranks as top AI model outside the US and China

Mistral has released Mistral Large 4, its largest model so far, as a public preview on its API. The company says the weights will follow by the end of October 2026. In independent tests by benchmarking firm Artificial Analysis, the trillion-parameter model scores 38 on the firm’s Intelligence Index. That makes it the highest-ranked model from outside the US and China, although several Chinese open-weight models score higher.

Mistral says Large 4 “pushes the frontier of open-weight performance.” The preview went live on October 6, 2026, through Mistral Studio. Mistral’s own announcement gives the model two nicknames: “unofficially ML4 and, very officially, le Chonk”. This article sets out what the model is, what it can do according to Mistral and independent testers, and how developers can reach it.

What Mistral Large 4 is

Large 4 is a mixture-of-experts (MoE) model. In this design, the network contains many specialised sub-networks, and only a few of them run for each token, so the compute per token is far smaller than the total size suggests. Mistral’s technical documentation gives 1.05 trillion total parameters, 52 billion active per token, and a separate 1.6 billion-parameter vision encoder.

The active-parameter figure varies by source. Artificial Analysis and VentureBeat reported 49 billion active parameters, while Mistral’s blog and docs say 52 billion. Mistral’s own research page lists the model as “1T - 49B active”. Tom’s Hardware gives the range as 49B–52B.

Mistral describes Large 4 as a hybrid “instruct-and-reasoning” model. It is meant to handle plain instruction following, multi-step reasoning and agentic work, meaning tool use and multi-step task automation, in a single model rather than separate products. Simon Willison reports that the API exposes two reasoning levels, none and high. Setting none switches reasoning off for a request.

Mistral also says Large 4 will be the base for a new generation of specialised Mistral models aimed at the industries its customers care about most.

Key specifications

Spec	Mistral Large 4 (preview)
Total parameters	1.05 trillion (Mistral docs)
Active parameters	52B (Mistral docs); 49B (Artificial Analysis, VentureBeat)
Vision encoder	1.6B parameters
Input / output	Text and images in, text out
Images per request	Up to 100, up from 8 on earlier Mistral models
Context window	1M tokens (Mistral docs); 512k measured in early preview
Languages	More than 160, including every official EU language
API model names	mistral-large-4, mistral-large-4-0

What is new compared with Mistral Large 3

The most concrete measured change is in document and image reasoning. On GDP.pdf, an Artificial Analysis test of reasoning over documents, Large 4 scores 19%. That is 18 points above Mistral Large 3 and level with Xiaomi’s MiMo-V2.6-Pro, though behind Moonshot’s Kimi K3 at 22%. Artificial Analysis attributes part of the gain to Mistral’s API, which now accepts 100 images per request instead of 8.

Languages are the other headline change. Mistral says the training data spanned more than 160 languages, and its research page describes the model as “natively fluent in 160+ languages”.

Mistral says it trained Large 4 from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own datacenters in Europe. The Next Web and VentureBeat reported about 4,000 GPUs, but Mistral’s announcement says 3,800. According to VentureBeat and The Next Web, training took roughly two months, and The Next Web reports it used about 10 megawatts. The model is described as the first milestone on the roadmap funded by Mistral’s 3 billion euro Series D, announced in September 2026.

Independent results from Artificial Analysis

Artificial Analysis gives Large 4 a score of 38 on version 4.3.2 of its Intelligence Index. The firm puts that alongside GPT-6 Luna at maximum effort (38) and DeepSeek V4.1 Flash at maximum effort (39). It calls Large 4 the “most intelligent model from outside the US and China,” ahead of models from countries such as South Korea and the United Arab Emirates.

The ranking has a clear limit. At least five Chinese open-weight models score higher on the same index:

Model	Index score	Total / active parameters	Cost per index task
Xiaomi MiMo-V2.6-Pro	46	1.0T / 42B	$0.13
Z.ai GLM-5.3 (max)	45	753B / 40B	$2.01
Moonshot Kimi K3 (max)	44	2.8T / 104B	$2.00
Z.ai GLM-5.3-Flash	42	320B / 18B	$0.25
DeepSeek V4.1 Flash (max)	39	552B / 16B	$0.27
Mistral Large 4 Preview	38	1.05T / 49B–52B	$1.13 ($0.57 at launch pricing)

Figures as compiled by Tom’s Hardware from Artificial Analysis data dated October 6.

Tom’s Hardware points out that MiMo-V2.6-Pro is about the same total size as Large 4. GLM-5.3-Flash, by contrast, is roughly one-third the size and scores four points higher.

Cybersecurity is the strongest result

Large 4 scores 50 on the Artificial Analysis Cyber Index. That is level with GLM-5.3-Flash and behind MiMo-V2.6-Pro at 56. Its best single result is on CyberGym-E2E-AA, an end-to-end test of reproducing and patching software vulnerabilities. There it scores 82%, ahead of MiMo-V2.6-Pro at 79% and GPT-6 Luna (max) at 78%. Artificial Analysis says that once the weights are released, Large 4 will rank among the top three open-weight models on its Cyber Index.

Mistral goes further and says the Cyber Index score places the model among the top five globally. Mistral also says Claude Opus 5.5 and GPT-6 Astra score near zero on CyberGym-E2E-AA because they refuse the task. That second point matters when reading the comparison: the gap partly reflects safety refusals by rival models, not only capability.

Mistral’s own coding and agent numbers

The coding and agent results below are Mistral’s figures, not independent measurements:

• DeepSWE v1.1: 61.7%
• SWE-Atlas-QnA: 59.4%
• Terminal-Bench 4: 28.3%
• Coding Agent Index: 49.8%
• AutomationBench: 59.9% across 657 multi-application workflows, ahead of Kimi K3 and DeepSeek V4 Pro by Mistral’s account

Those numbers put Large 4 ahead of DeepSeek V4 Pro and Qwen 3.8 Max, but behind the leaders on coding. VentureBeat noted that the live DeepSWE leaderboard shows GLM-5.3 and Kimi K3 at about 69%. It shows GPT-6 Astra, Gemini 3.8 Flash and Claude Opus 5 at about 74%.

In blind human coding evaluations Mistral ran with Surge AI on a 1-to-5 scale, Large 4 scored 3.74. That trails Claude Opus 5 at 4.22 and leads GLM-5.3 at 3.60. Mistral’s research page also claims state-of-the-art results among open-weight models on finance and manufacturing. The sources give no independent figures for those claims.

Access and pricing

Large 4 is available as a Research Public Preview on Mistral’s API, through Mistral Studio, under the model names mistral-large-4 and mistral-large-4-0. Mistral lists the following prices:

• Input: $1.36 per million tokens
• Output: $4.18 per million tokens
• Cached input: $0.14 per million tokens
• Launch discount: 50% off for the first two weeks, which brings input and output to $0.68 and $2.09 per million tokens

Price is a real factor in the comparison. Artificial Analysis calculates that Large 4 costs $1.13 per Intelligence Index task at standard pricing. That is over four times the cost of open-weight models with similar scores: GLM-5.3-Flash costs $0.25 per task and DeepSeek V4.1 Flash (max) $0.27. Even at the launch discount of $0.57 per task, Large 4 remains more expensive than both.

The open-weights promise

Open weights let teams download a model and run or fine-tune it on their own hardware. Mistral said Large 4’s weights are due “by the end of the month”. Independent reporting cited October 27, 2026 as the date. Mistral’s research page already lists Large 4 as an “open-weight” model. However, none of the sources confirms that the weights have been published, and none states the licence terms.

What it means for builders

For teams that want a European-built model, Large 4 is now Mistral’s strongest option. Mistral trained it in its own European datacenters, and Artificial Analysis ranks it the top model from outside the US and China. Three features stand out for practical work:

• The language coverage, which includes every official EU language.
• The higher image limit per request, useful for document-heavy pipelines.
• The ability to switch reasoning off per request when speed matters more than depth.

Security tooling is the clearest fit on current evidence. The CyberGym-E2E-AA result is independent, and it beats rival models that refuse the task.

There are also limits to weigh:

• Cost per task: On general intelligence, Chinese open models score higher at a fraction of the cost per task.
• Context window: The documented 1M tokens differs from the 512k Artificial Analysis observed in the early preview. Developers relying on long context should test the limit themselves.
• Self-reported results: Most coding and agent figures come from Mistral.
• Preview status: The model is labelled a research preview, so behaviour and pricing may change before a general release.

What to watch next

The main open question is the weights: whether they arrive on schedule, under what licence, and whether the released version matches the preview that Artificial Analysis tested. Once the weights are out, independent providers can host the model, which may change the cost comparison with Chinese open models.

Watch, too, for the specialised industry models Mistral says it will build on top of Large 4, and for independent checks of its coding and agent claims. Developers building agents on Mistral’s platform may also want to read Unicorn API’s earlier coverage of the Mistral Responses API preview.