🔍 Read the full analysis: A Closer Look At Mistral Large 4 And The AI Frontier on ThorstenMeyerAI.com

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

Get the latest gadgets 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 launched Mistral Large 4 in public API preview on October 6, adding its largest model yet to a crowded frontier AI market. An October 7 report by Thorsten Meyer cites an Artificial Analysis Intelligence Index score of 38 and argues that the current preview is not a preferred choice for demanding, long-running agentic tasks; that assessment combines benchmark data with the author’s own experience, not a controlled reliability study.

Mistral describes Large 4 as a mixture-of-experts model with one trillion total parameters and 49 billion active parameters. The preview accepts text and images through an API. The company says it trained the model on its own infrastructure in Europe and is continuing to improve it. Mistral has said the model weights are scheduled for release later in October, but they were not publicly downloadable as of the report’s October 7 publication.

Artificial Analysis’s score of 38 on its Intelligence Index matches OpenAI’s GPT-6 Luna at maximum reasoning effort and is one point below DeepSeek V4.1 Flash at maximum effort. In the same dated snapshot, the index lists Anthropic’s Claude Opus 5.5 at 58, Google’s Gemini 4 Argon at 53, OpenAI’s GPT-6.1 Sol at 52, Z.ai’s GLM-5.3 at 45 and Moonshot AI’s Kimi K3 at 44. These are benchmark scores under the settings named by the source, not evaluations made under identical compute budgets.

Meyer says that his own use of the preview produced hallucinations and lowered his confidence in assigning it longer tasks. He presents this as personal experience, not a controlled comparison of hallucination rates. The source also notes that the index’s aggregate score does not directly measure reliability on a particular coding, research or tool-using workflow.

How Large 4 Stacks Up

The launch gives developers another model to test, but the available evidence does not establish that Large 4 matches the strongest models for complex work. In the October 7 Artificial Analysis snapshot, the gap between Mistral’s score of 38 and the listed leading U.S. models ranges from 14 to 20 index points. The source cautions that these are index-point differences, not percentages or direct predictions of task success.

That distinction matters for agentic work, where a model may plan, call tools and carry decisions through several steps. Errors or unsupported assumptions early in a process can affect later actions. A benchmark can inform model selection, but it cannot by itself show how often a system follows constraints, checks evidence or recovers from mistakes in a particular workflow.

Mistral’s release also has significance for European AI capacity: the company says it trained the model on its own European infrastructure. That is a separate point from whether the preview is the best choice for a given developer. The source’s comparison includes a counterexample to any claim that every rival scores higher: Cohere Command A+, a Canadian model, received 13 on the same index snapshot. Mistral’s score is higher, but that does not close the gap with the leading U.S. and stronger Chinese entries listed.

Preview First, Weights Later

The key status distinction is that Large 4 is currently a preview API, not a completed public release of downloadable weights. Mistral announced the preview on October 6. The source report, published October 7, says the weights are planned for later in the month and that Mistral is still working on improvements. Those future steps could change what developers can access and how the model performs, but they were not yet available at the time described.

Artificial Analysis reports a context capacity of about 512,000 tokens. That figure describes how much input material the model can accept; it does not establish that the model will reason accurately across all of it. Likewise, Mistral’s advertised strengths in agentic coding and specialized professional tasks are company claims that require testing on the workloads developers intend to run.

The index comparison is a dated snapshot from October 7, 2026, and scores may change. The developer locations in the table refer to where the companies are based, not where a specific API request is processed. The source says DeepSeek V4.1 Flash has approximately comparable benchmark intelligence to Large 4 at a much lower measured cost per task, but the supplied material does not give the underlying price figures or a full cost methodology.

“I would not choose it for demanding agentic work or long tasks when stronger models are available.”

— Thorsten Meyer, author of the October 7 report

What the Preview Cannot Yet Show

The supplied material does not establish how Large 4 performs across a broad, independently controlled set of real-world agentic tasks. The Intelligence Index is an aggregate benchmark, while Meyer’s report of hallucinations reflects his own use. Neither proves how the model will perform on a specific developer’s workload or how its error rate compares with competitors under matched conditions.

It is also unclear from the source when the promised weights will become available beyond the stated plan for later in October, what changes Mistral may make during the preview, and whether those changes will affect benchmark results. The source refers to a cost comparison with DeepSeek but does not provide the figures or enough methodology to assess its applicability to different usage patterns. API processing locations are not established by the developer-location comparison.

Watch for Weights and Retests

The next stated milestone is Mistral’s planned release of Large 4 model weights later in October. Mistral also says it is continuing to improve the model. Once updates or weights arrive, developers and independent evaluators can check whether benchmark scores change and test the system on specific coding, research and tool-use tasks.

For now, the practical next step for teams considering the preview is to evaluate it against their own requirements rather than treating the index score or the model’s parameter count as a guarantee of performance. Results on sustained task completion, factual support, oversight needs and cost would help clarify where Large 4 is useful. The source does not provide a date for a final release or a schedule for further independent evaluations.

Key Questions

What did Mistral announce?

Mistral introduced Mistral Large 4 as a public API preview on October 6, 2026. The model accepts text and images and is described as having one trillion total parameters, with 49 billion active parameters.

Are Large 4’s weights publicly available?

Not according to the October 7 report. Mistral had scheduled a release of the weights for later in October, but they were not publicly downloadable at the time of publication.

How did Large 4 score on the cited benchmark?

Artificial Analysis gave the preview an Intelligence Index score of 38 in the snapshot reported October 7, 2026. The source says that matches GPT-6 Luna at maximum reasoning effort and is just below DeepSeek V4.1 Flash at maximum effort. Scores and settings may change, and the comparison is not under identical compute budgets.

Does the score prove Large 4 is unreliable?

No. The index is an aggregate benchmark, not a direct reliability test for every task. Meyer reports hallucinations from personal use, but says that experience is not a controlled comparative study. Developers would need to test the preview on their own workflows to judge its performance.

What should developers watch next?

The main stated milestone is the planned release of model weights later in October, alongside further model improvements. New benchmark results and workload-specific evaluations could clarify how the model performs as its preview develops.

Source: ThorstenMeyerAI.com