🔍 Read the full analysis: Mistral Large 4 Is Still Chasing The AI Frontier on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Mistral released Large 4 as an API preview on October 6, 2026. Artificial Analysis gave it an Intelligence Index score of 38, below several leading US and Chinese models; the model’s weights have not yet been released. A reviewer advises against using the current preview for demanding agentic tasks, based partly on personal experience that is not a controlled comparison.
Mistral AI launched its Mistral Large 4 model in public API preview on October 6, but an October 7 benchmark snapshot places it behind several leading US and Chinese models, raising questions about its suitability for demanding agentic work. The model’s weights are expected later in October and are not yet publicly downloadable.
Mistral describes Large 4 as a mixture-of-experts model with one trillion total parameters and 49 billion active parameters. It accepts text and images. The company says it trained the model on its own infrastructure in Europe and is continuing to improve it. These specifications and statements come from Mistral’s announcement; they do not, on their own, establish how reliably the preview performs on particular tasks.
Artificial Analysis assigned the preview an Intelligence Index score of 38 in data available on October 7. That matched OpenAI’s GPT-6 Luna at maximum reasoning effort and was just below DeepSeek V4.1 Flash at maximum effort, which scored 39. The same snapshot listed Claude Opus 5.5 at 58, Gemini 4 Argon at 53, GPT-6.1 Sol at 52, GLM-5.3 at 45 and Kimi K3 at 44. These are index points, not percentages, and the evaluations use differing model settings rather than identical compute budgets.
Thorsten Meyer, writing on ThorstenMeyerAI.com on October 7, said he would not select the current preview for long or demanding agentic tasks when stronger alternatives are available. He cited the benchmark results and his own use of the model, during which he said he encountered hallucinations. That account is personal experience, not a controlled comparison; it does not establish hallucination rates across models or show that Large 4 will fail a specific task. Meyer also noted that Artificial Analysis reports a context capacity of roughly 512,000 tokens, while cautioning that capacity does not guarantee accurate reasoning across a long input.
Model Watch · October 7, 2026
Mistral Large 4 Is Still Chasing the AI Frontier
Mistral’s new API preview brings a trillion-parameter mixture-of-experts model into public testing. An October 7 benchmark snapshot puts it behind several leading alternatives, while its fit for demanding agentic work remains unsettled.
01 / Benchmark snapshot
A visible gap to the leaders
Artificial Analysis listed Large 4 at 38 on October 7. Several cited US and Chinese models scored higher. Settings differed across evaluations, so the index is useful context rather than a controlled, equal-compute contest.
Mixture of experts
Mistral describes one trillion total parameters, with 49 billion active parameters. It accepts text and images.
API preview only
The public preview opened October 6. Weights were not downloadable as of October 7; Mistral said they were planned for later in October.
Test on real work
Mistral cites agentic coding and specialized professional uses. Those claims need workload-specific evaluation before deployment decisions.
02 / Choosing a model
Why the benchmark gap matters
For coding, research, and workflows that call tools across many steps, an early mistake can shape later decisions. A fluent answer alone cannot show whether the process was reliable. Yet an aggregate index cannot determine the outcome of a particular job.
Define the workload
Use representative coding, research, or tool-use tasks.
Track quality
Check completion, unsupported claims, and multi-step errors.
Measure oversight
Record supervision needs, retries, and human correction.
Compare total cost
Include task cost and the effort needed to reach a sound result.
03 / Evidence and limits
What the preview can—and cannot—tell us
Thorsten Meyer advised against choosing the current preview for long or demanding agentic tasks when stronger alternatives are available. He cited the index and his own use, during which he reported hallucinations. That is personal experience, not a controlled comparison or a measured cross-model rate.
What is supported
The dated index places Large 4 below several cited alternatives. Mistral opened API access and says the model was trained on its own European infrastructure.
What remains unknown
The supplied evidence does not establish long-task reliability, comparative hallucination rates, or how performance may change as Mistral continues improving the preview.
Capacity is not accuracy
A reported context capacity of roughly 512,000 tokens describes how much input can fit. It does not guarantee sound reasoning across a long input.
Cost comparison is incomplete
The source says DeepSeek V4.1 Flash has comparable benchmark intelligence at lower measured cost per task, but provides no prices or detailed methodology.
04 / Voices & context
Three useful cautions
“The model was trained on Mistral’s own infrastructure in Europe, and the company says it is continuing to improve it.”
“I would not choose it for demanding agentic work or long tasks when stronger models are available.”
“A score of 38 does not prove that Mistral will fail a particular coding or research task.”
Training infrastructure in Europe is relevant to Mistral’s capacity-building effort. It does not establish where every API request is processed. The benchmark is an October 7 snapshot and may change as models and evaluations are updated.
05 / What comes next
Weights and workload tests ahead
Mistral said weights were planned for later in October 2026 and that it was continuing to improve Large 4. The supplied information does not confirm a release date beyond that window. Until then, the clearest picture is limited: a public preview, an index score below several cited alternatives, and open questions about complex autonomous work.
What did Mistral release?
A public API preview on October 6, 2026. The model accepts text and images and is described as a mixture-of-experts model.
Can the weights be downloaded?
Not as of October 7 in the source material. Mistral said weights were scheduled for later in October.
How did it score?
Artificial Analysis assigned 38 index points. Claude Opus 5.5 scored 58, GLM-5.3 scored 45, and DeepSeek V4.1 Flash scored 39.
What should teams evaluate?
Test the preview on their own tasks, tracking output quality, unsupported claims, supervision needs, and cost.
Benchmark Gap Shapes Model Choice
The launch gives developers another model to test, but the current benchmark snapshot offers limited evidence of parity with leading systems. For companies considering models for coding, research or workflows that use tools over many steps, the distinction matters: a mistaken assumption early in a task can affect later decisions, and a fluent final answer does not prove that the process was sound.
Still, an aggregate index is not a direct reliability test for every workload. A score of 38 does not predict whether Large 4 will succeed or fail on a particular coding job, and the reported evaluation settings are not identical. Developers should test the preview against their own tasks and error costs rather than treat the ranking as a definitive verdict. Mistral’s claims about agentic coding and specialized professional work also require workload-specific assessment.
The comparison has limits in both directions. The snapshot places Cohere’s Command A+ at 13, below Mistral, so it would be inaccurate to say every major competitor scored higher. The supported point is narrower: several leading US models and stronger Chinese alternatives scored above Large 4 on this index. The figures do not represent percentage differences in intelligence or guaranteed differences in task success.
Products Worth Considering
Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Black
Apple iPad Air (2025) 13‑inch Wi‑Fi 256 GB – Space Gray (M3 chip)
Samsung Galaxy Tab S10 FE 128GB Wi-Fi Android Tablet, Blue
As an affiliate, we earn on qualifying purchases.
A Preview Before Weight Release
The timing and distribution model matter when interpreting the launch. On October 6, Mistral made a preview API available; it did not release downloadable weights at that point. The company said weights were scheduled for later in October. That planned release could change how developers access or adapt the model, but the current evaluation concerns the API preview, not a publicly downloadable version.
Mistral says Large 4 was trained on its own European infrastructure. That is relevant to the company’s effort to build European AI capacity, but the location of training does not establish where every API request is processed. Likewise, developer location in the benchmark comparison identifies the organizations behind the models, not the location of a particular user’s request.
Artificial Analysis’s figures are a dated snapshot from October 7, 2026, and may change as models and evaluations are updated. The source material also says DeepSeek V4.1 Flash has approximately comparable benchmark intelligence at a much lower measured cost per task, but it provides no prices or detailed cost methodology here. That cost comparison should not be treated as a complete assessment of total deployment expense.
“The model was trained on Mistral’s own infrastructure in Europe, and the company says it is continuing to improve it.”
— Mistral AI, in its October 6 announcement
Preview Limits and Open Questions
The evidence available here does not establish how Large 4 performs across a broad, controlled set of long-running agentic tasks, nor does it provide a comparative hallucination rate. Meyer’s account of hallucinations reflects his own use; the source does not describe a controlled study, task sample or measurement method. The benchmark score also cannot settle how the model will perform on a developer’s specific workload.
The model’s weights had not been released as of October 7, and their eventual availability and form remained pending. It is also not clear from the supplied information how the preview’s performance, pricing or reliability may change as Mistral continues its work. The reported lower cost for DeepSeek V4.1 Flash is not accompanied here by figures or methodology, so its practical cost advantage cannot be independently assessed from this material.
Weights and Workload Tests Ahead
Mistral said it planned to release Large 4’s weights later in October 2026 and was continuing to improve the model. Those are company-stated plans; the supplied information does not confirm a release date beyond that window. Developers and evaluators can compare the API preview with their own coding, research and tool-use tasks, tracking not only completion quality but also unsupported claims, supervision needs and cost.
Further benchmark updates may change the dated ranking. Until then, the clearest assessment is limited: Large 4 is available as a preview, its index score trails several cited alternatives, and its suitability for complex autonomous work remains a question for task-specific testing.
Key Questions
What did Mistral release?
Mistral made Large 4 available through a public API preview on October 6, 2026. It accepts text and images and is described as a mixture-of-experts model with one trillion total parameters and 49 billion active parameters.
Are Mistral Large 4’s weights available to download?
Not as of October 7, 2026. Mistral said weights were scheduled for release later in October, but that future release had not yet taken place in the source material.
How did Large 4 score against other models?
Artificial Analysis gave the preview an Intelligence Index score of 38 in its October 7 snapshot. Several cited models scored higher, including Claude Opus 5.5 at 58 and GLM-5.3 at 45. The scores are index points, and evaluations used different reasoning settings.
Does the benchmark prove Large 4 is unreliable?
No. An aggregate benchmark score does not establish performance on every task or provide a direct reliability rate. The reviewer’s comments about hallucinations are personal observations, not results from a controlled comparative study.
What should developers watch for next?
Watch for the planned weight release later in October, further benchmark results and independent tests on real coding, research and agentic workflows. The date and eventual performance remain subject to updates.
Source: ThorstenMeyerAI.com
