🔍 Read the full analysis: Mistral Large 4 Remains Behind The AI Frontier on ThorstenMeyerAI.com
Get school and study 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
TL;DR
Mistral launched Large 4 in public API preview on October 6, 2026. Artificial Analysis gives it an Intelligence Index score of 38, below several leading U.S. and Chinese models; the source article’s recommendation against using it for demanding agentic work is an assessment, not a confirmed measure of performance on every task.
Mistral launched Large 4 in public API preview on October 6, but an Artificial Analysis benchmark snapshot published the following day scored it 38 on the Intelligence Index, below several leading U.S. and Chinese models. The release is a notable step for Mistral’s model lineup, but the available evidence does not establish that the preview matches those competitors for demanding work.
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. Mistral said it trained the model on its own infrastructure in Europe and is continuing to improve it. The company scheduled a release of the model weights for later in October; they were not publicly downloadable as of the October 7 report.
In the Artificial Analysis comparison reported by Thorsten Meyer, Large 4 Preview scored 38 index points. The same snapshot placed Claude Opus 5.5 at 58, Gemini 4 Argon at 53 and GPT-6.1 Sol at 52. Chinese models GLM-5.3 and Kimi K3 scored 45 and 44, respectively, while DeepSeek V4.1 Flash scored 39. These are benchmark index points, not percentages or direct predictions of task success.
The comparison is not a controlled test at identical reasoning or compute settings: several competing scores use named maximum or high reasoning settings, while the listed Large 4 score is for its preview. Meyer also reports that his own use of the preview produced hallucinations that reduced his confidence in assigning it long tasks. That is personal experience, not a controlled comparison of hallucination rates.
Model report · October 7, 2026
Mistral Large 4 Remains Behind The AI Frontier
Mistral’s public API preview is a major addition to its lineup. A dated benchmark snapshot places it behind several leading U.S. and Chinese models, while leaving important real-world questions open.
At a glance
A notable launch, with a measured gap
Mistral says it trained Large 4 on its own infrastructure in Europe and continues to improve it. The October 7 report describes an API preview; model weights were scheduled for later in October and were not yet downloadable at that time.
The reported benchmark offers one comparison point. It does not establish performance on every task, nor does a long context window guarantee reliable reasoning across all the material it can hold.
Benchmark snapshot
Large 4 scored 38 index points
Artificial Analysis figures reported October 7, 2026. These are index points, not percentages or direct predictions of task success.
Large 4 is close to DeepSeek V4.1 Flash and above Command A+ in this reported comparison.
Settings were not identical: several competitors used named high or maximum reasoning modes, while the listed Large 4 score was for its preview. This is not a controlled head-to-head test.
A reason to evaluate carefully
The snapshot puts Large 4 near some comparison models and below several leaders. It can inform a shortlist, but it cannot determine how a model will perform on a specific coding, research, or tool-use workload.
Errors can compound
Long tool-using sequences depend on sound planning and decisions at each step. A fluent final answer alone does not show that the process was reliable.
Not a controlled test
Thorsten Meyer reports hallucinations during his own use and lower confidence for long tasks. This is individual experience, not a comparative hallucination study.
A broader milestone
Mistral says it trained Large 4 on its own European infrastructure. The benchmark does not settle deployment cost, privacy, reliability, or request-processing location.
Release timeline
Preview now; weights were planned
The October 7 report records what was available and what Mistral said it expected next.
Preview announced
Mistral introduces Large 4 for public API access with text and image inputs.
Snapshot published
Artificial Analysis score of 38 is reported alongside selected model results.
Weights expected
Mistral scheduled a weights release; the report does not confirm it happened.
Test real workloads
Compare task quality, long-sequence reliability, supervision, and total operating cost.
Evidence boundaries
What this snapshot cannot show
The available record supports a dated benchmark comparison and an attributed individual assessment. It does not give a definitive verdict for every use case.
Whether the weights arrived later in October, how much the preview changed, and how Large 4 performs across independent professional workloads.
Useful next checks for developers
Key questions
Four points to keep in view
What did Mistral announce?
A public API preview on October 6, 2026: a mixture-of-experts model described as one trillion total parameters and 49 billion active parameters, with text and image input.
How did Large 4 score?
Artificial Analysis gave the preview 38 Intelligence Index points in the October 7 snapshot. It is an index result, not a percentage or task guarantee.
Are the weights available?
Not according to the October 7 report. Mistral said weights were planned for later in October; the source does not confirm their release.
Does the benchmark prove it is unreliable?
No. An aggregate score and one author’s experience cannot establish reliability or hallucination rates across all agentic tasks.
What the Benchmark Gap Signals
The snapshot gives developers a reason to distinguish a major model release from demonstrated frontier performance. A score of 38 is close to DeepSeek V4.1 Flash at 39 and matches GPT-6 Luna at maximum reasoning effort in the reported comparison, but it trails several other models that developers may consider for complex work. The index is useful as one comparison point; it cannot establish how a model will perform on a particular company’s coding, research or tool-use tasks.
That distinction matters for agentic workflows, where a system plans, uses tools and carries decisions across multiple steps. Errors or unsupported assumptions can compound, and a fluent final response does not by itself show that the steps were sound. Meyer argues that he would choose stronger alternatives for demanding, extended tasks. This is his evaluation of the preview, not a finding that Large 4 fails all such work.
The launch also has significance beyond rankings. Mistral says it trained the model using its own European infrastructure, and the company is presenting it as a substantial addition to its product line. That is relevant to European AI capacity. But the benchmark snapshot does not settle questions about deployment costs, reliability, privacy or performance under specific workloads.
As an affiliate, we earn on qualifying purchases.
Preview, Weights and Comparisons
The report distinguishes the current API preview from the planned weights release. As of October 7, developers could access the preview through Mistral’s API, but could not download the weights. Mistral said the weights were expected later in October and that model improvements were ongoing; the report does not confirm that either development had already happened.
The scores are a dated snapshot from October 7, 2026, attributed to Artificial Analysis. The source lists each competitor’s developer location and evaluated reasoning setting. Those settings are not identical, so the figures should not be read as a definitive head-to-head result under matched conditions. Developer location also does not show where an individual API request is processed.
Meyer’s comparison includes Cohere Command A+ at 13, a lower score than Large 4 on this index. That is a counterexample to a claim that every competitor scores higher, but it does not remove the gap with the higher-ranked U.S. and Chinese models. The source also notes Artificial Analysis’s estimate of roughly 512,000 tokens of context capacity for Large 4. Context length describes how much material a request can contain; it does not, by itself, demonstrate reliable reasoning across that material.
“Mistral said it trained the model on its own infrastructure in Europe and continues to improve it.”
— Mistral, as described in its October 6 announcement
AI model performance testing software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What the Snapshot Cannot Show
The Intelligence Index is an aggregate benchmark, and the reported comparison does not use identical reasoning settings or compute budgets for every model. The figures therefore cannot establish which system will be more reliable on a particular developer’s tasks, or quantify how often Large 4 hallucinates compared with its rivals. Meyer’s account of hallucinations is based on his own use, not a controlled study.
It is also unclear from the source how much the preview may change before the planned weights release, whether the weights arrived later in October, and how the model performs across independently tested coding, research and professional workloads. The report mentions a cost comparison but its supplied text ends before giving the figures, so it does not establish a cost advantage or disadvantage for Large 4.
As an affiliate, we earn on qualifying purchases.
Weights Release and Further Testing
Mistral said the model weights were scheduled for release later in October 2026 and that it was continuing to improve Large 4. Those were forward-looking statements in the October 7 report, not completed milestones. A later status update would be needed to confirm whether the weights were released and whether the preview changed.
For developers weighing adoption, the next useful evidence would be workload-specific evaluations: performance on their own tasks, reliability over long tool-using sequences, supervision demands and total operating cost. Further benchmark results using clearly matched settings could also make comparisons more informative. Until then, the available record supports a dated benchmark comparison and an attributed individual assessment, not a definitive verdict on every use case.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did Mistral announce?
Mistral announced Large 4 in public API preview on October 6, 2026. It described the model as a mixture-of-experts system with one trillion total parameters and 49 billion active parameters that accepts text and images.
How did Mistral Large 4 score?
Artificial Analysis gave the preview an Intelligence Index score of 38 in the snapshot reported on October 7. The score is an index result, not a percentage or a guarantee of performance on an individual task.
Are the model weights publicly available?
Not according to the October 7 report. Mistral said the weights were scheduled for release later in October, but the source does not confirm that release occurred.
Does the benchmark prove Large 4 is unreliable for agentic work?
No. The score is an aggregate benchmark result, not a direct test of every agentic workflow. Meyer recommends stronger alternatives for demanding, extended tasks, but presents that as his judgment; his hallucination observations are personal experience, not a controlled study.
What should developers evaluate before using it?
Developers can test Large 4 on their own coding, research or tool-use workloads, including long sequences that require checking evidence and following constraints. The source does not provide enough cost data to determine whether the preview is cheaper or more expensive than competing models.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
