🔍 Read the full analysis: The AI Frontier Gap Mistral Large 4 Has Yet To Close on ThorstenMeyerAI.com

Prime Big Deal Days · Oct 6–7Offer from Amazon

Get office and shipping supplies delivered free — and shop member deals

• Fast, free delivery on millions of items
• Access to Prime Big Deal Days deals on October 6–7
• Prime Video, Amazon Music and more included

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

Mistral AI released Mistral Large 4 in public API preview on October 6, but an early benchmark snapshot places it behind leading U.S. models and several Chinese competitors. Artificial Analysis gave the model a score of 38 on its Intelligence Index, prompting the source author to say they would not choose the preview for demanding agentic work or long tasks when higher-scoring options are available.

Mistral describes Large 4 as its largest model to date: a mixture-of-experts system with one trillion total parameters and 49 billion active parameters. The preview accepts text and images. Mistral said it trained the model on its own infrastructure in Europe and is continuing to improve it. The company plans to release the model weights later in October, but they were not publicly downloadable as of the source report on October 7.

In the Artificial Analysis comparison published for that date, Mistral Large 4 Preview scored 38 index points. Claude Opus 5.5, evaluated at maximum reasoning effort with default fallback, scored 58; Gemini 4 Argon at high effort scored 53; and GPT-6.1 Sol at maximum effort scored 52. Chinese models GLM-5.3 and Kimi K3 scored 45 and 44, respectively, while DeepSeek V4.1 Flash at maximum effort scored 39. GPT-6 Luna at maximum effort also scored 38.

The comparison is a dated snapshot, and the reasoning settings differ, so it is not a like-for-like test under identical compute budgets. The source report says the numbers are index points, not percentages or direct predictions of task success. It also notes that developer location does not establish where a specific API request is processed. Cohere’s Command A+ scored 13, a counterexample to any claim that every named competitor outperformed Mistral.

Benchmark Gap Shapes Model Choice

The score matters most as an early signal for developers deciding which model to test or deploy. A system used for agentic work may need to plan, call tools, interpret results and carry decisions through several steps. Errors or unsupported assumptions early in that chain can affect later actions, even if the final response sounds coherent. A lower aggregate score does not prove failure on a particular workflow, but it gives buyers a reason to compare performance against their own tasks before relying on the preview for complex work.

The benchmark does not settle whether Large 4 is useful, nor whether it is the best choice for a specific task, cost target or deployment requirement. Mistral advertises strengths in agentic coding and specialized professional work, according to the source report; those claims require workload-specific evaluation. The report’s author says their own experience with hallucinations weakened their confidence, while explicitly describing that experience as anecdotal rather than a controlled comparison.

For European AI capacity, the launch also marks a significant product development: Mistral says the model was trained on its own European infrastructure. But infrastructure and scale do not by themselves demonstrate parity with benchmark leaders. The distinction between a model’s technical specifications and its measured performance is relevant to companies weighing alternatives for long-running or high-stakes workflows.

Preview First, Weights Later

The development is an API preview, not a completed public release of downloadable model weights. Mistral announced Large 4 on October 6. As of the following day, the weights were scheduled for release later in October, while the preview could be accessed through an API. That timing matters for developers who may be evaluating model access, deployment control or the ability to run a system themselves.

Artificial Analysis reports a context capacity of roughly 512,000 tokens. A context window describes how much material a model can receive in a request; it does not establish that the model will reason reliably across all of it. Similarly, the one-trillion total parameter count and 49-billion active parameter count describe the model’s architecture, not a guarantee of task quality.

The scores cited here come from the Intelligence Index snapshot available on October 7, 2026, as reported by ThorstenMeyerAI.com and attributed there to Artificial Analysis. Scores may change as models or evaluations are updated. The report also says DeepSeek V4.1 Flash offers approximately comparable benchmark intelligence at a much lower measured cost per task, though the supplied material does not give the underlying cost figures or further testing details.

“The model was trained on Mistral’s own infrastructure in Europe, and the company says it is continuing to improve it.”

— Mistral AI, as described in its announcement

Performance Beyond the Index

The available score is an aggregate benchmark result, not a direct test of every coding, research or business workflow. It does not show how often Mistral Large 4 will make errors on a particular task, how reliably it will use tools over many steps, or how it compares under identical reasoning and compute settings. The author’s report of hallucinations is based on personal experience; the supplied material does not provide a controlled head-to-head hallucination study.

Further details also remain unavailable in the source material. It gives no underlying figures for the reported cost comparison with DeepSeek V4.1 Flash, and it does not establish whether the planned weight release occurred after the report’s October 7 publication. Mistral’s preview may improve, but the scale and timing of any performance changes are not specified.

Weights and Workload Tests Ahead

The next stated milestone is Mistral’s planned release of Large 4’s model weights later in October 2026. The source report does not confirm a specific release date or say whether that release will proceed as planned. Updated benchmark results and independent testing could also change the comparison as the preview develops.

For developers, the practical next step is to test the model on representative tasks and compare it with alternatives using the same prompts, tools and evaluation criteria. Longer agentic runs, factual verification, error rates and total cost would help clarify whether the preview suits a particular use case. Until that evidence is available, the current benchmark snapshot supports caution about treating Large 4 as a proven choice for demanding autonomous work.

Key Questions

What has Mistral released?

Mistral introduced Mistral Large 4 as a public API preview on October 6, 2026. The model accepts text and images. Its weights were not publicly downloadable as of the October 7 report.

How did Mistral Large 4 score?

Artificial Analysis gave Mistral Large 4 Preview a score of 38 on its Intelligence Index in the snapshot cited on October 7, 2026. The score is an aggregate benchmark result, not a percentage or a direct forecast for every task.

Does the score prove Mistral Large 4 will fail at agentic work?

No. The score does not establish how the model will perform on a specific workflow. It is one data point that can guide evaluation; developers should test the preview on their own tasks, including tool use and longer sequences of decisions.

When are the model weights expected?

Mistral said the weights were scheduled for release later in October 2026. The source report, dated October 7, does not confirm a specific date or whether the release subsequently happened.

Source: ThorstenMeyerAI.com