🔍 Read the full analysis: Mistral Large 4’S Progress Hasn’t Caught The AI Frontier on ThorstenMeyerAI.com
Get tech for your team 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 as an API preview on October 6, with a one-trillion-parameter mixture-of-experts architecture and planned weight release later in October. Artificial Analysis gave it an Intelligence Index score of 38, below several leading U.S. and Chinese models; the score is a dated benchmark snapshot, not a direct measure of success on every task.
Mistral AI launched Mistral Large 4 in public API preview on October 6, but an October 7 benchmark snapshot places it behind several leading U.S. and Chinese models. The preview scored 38 on Artificial Analysis’s Intelligence Index, leaving its standing for demanding agentic tasks an open question rather than establishing it as a frontier leader.
Large 4 is a mixture-of-experts model with one trillion total parameters and 49 billion active parameters, according to Mistral’s announcement. It accepts text and images. The company said it trained the model on its own infrastructure in Europe and is continuing to improve it. The model is currently available through a preview API; Mistral scheduled the release of its weights for later in October, so they were not yet publicly downloadable at the time of the source report.
Artificial Analysis’s Intelligence Index comparison, dated October 7, gives Large 4 Preview a score of 38. That matches OpenAI’s GPT-6 Luna at maximum reasoning effort and is just below DeepSeek V4.1 Flash at maximum effort, which scored 39. The same snapshot lists Z.ai’s GLM-5.3 at 45 and Moonshot AI’s Kimi K3 at 44. U.S. models score higher: Anthropic’s Claude Opus 5.5 at 58, Google’s Gemini 4 Argon at 53 and OpenAI’s GPT-6.1 Sol at 52.
These are benchmark index points, not percentages or predictions of success on a particular task. The comparison also uses different named reasoning settings rather than identical compute budgets. Artificial Analysis reports a context capacity of roughly 512,000 tokens, but that measures how much material can fit into a request; it does not establish how reliably the model reasons over that material.
AI Frontier · Benchmark Snapshot · October 7, 2026
Mistral Large 4’s Progress Hasn’t Caught the AI Frontier
A trillion-parameter model has entered public API preview. Its first benchmark snapshot trails several leading U.S. and Chinese models, while real-world performance remains a workload-specific question.
01 / The scorecard
A clear gap in this snapshot
Selected Intelligence Index results published October 7, 2026. Bars show index points.
For context, Cohere Command A+ scored 13 in the same snapshot. The comparison uses different named reasoning settings, not identical compute budgets; scores do not predict success on a particular task.
02 / What the evidence says
Promising scale. Open questions.
The launch details and early results answer different questions. Neither parameter count nor context size tells the whole story.
Large by design
Mistral describes a text-and-image mixture-of-experts model with one trillion total parameters and 49 billion active parameters.
Preview first
Public API access began October 6. Weights were scheduled for later in October, so developers could not yet download and run them at the time assessed.
Capacity is not reliability
A roughly 512,000-token context can fit substantial material in a request. It does not establish how accurately the model reasons across it.
03 / Demanding work
Why agentic tasks need more than a score
Multi-step work can compound early mistakes. A polished final answer alone does not show that each step was sound.
Plan
Break the goal into steps and keep assumptions clear.
Use tools
Act on systems, inspect outputs, and recover from errors.
Carry context
Use earlier findings consistently across a long workflow.
Verify
Check accuracy and unsupported claims in the completed work.
04 / Reviewer perspective
A caution, clearly bounded
The source report combines a benchmark snapshot with one reviewer’s own experience.
“I would not choose it for demanding agentic work or long tasks when stronger models are available.”Thorsten Meyer · ThorstenMeyerAI.com
Experience, not a controlled trial
Meyer reported encountering hallucinations during his own use and said this reduced his confidence in assigning longer tasks. That is a personal observation; it does not establish hallucination rates or show that competitors never make unsupported claims. Mistral’s claims about agentic coding and professional work still need workload-specific evaluation.
05 / Limits of the snapshot
What it cannot establish
Read the October 7 index as one comparative signal, with several boundaries.
Individual workloads vary
The score does not show how Large 4 performs on a particular team’s coding, research, or business tasks.
Settings differ
Named reasoning settings and compute budgets are not identical across the compared models. No controlled side-by-side test of hallucinations or long-task completion is provided.
Price and reliability remain open
The excerpt does not give a full cost comparison. The benchmark alone cannot establish product value, reliability, or suitability.
06 / Next steps
Test the preview on real work
The planned weight release would enable another kind of evaluation; the source material does not confirm its eventual timing or license terms.
Evaluate complete workflows
Try the available preview on representative tasks. Track accuracy, unsupported outputs, tool use, and performance from start to finish.
Keep the date attached
Mistral said it was continuing to improve Large 4. Future benchmarks may change the comparison; this snapshot describes the preview as assessed on October 7, 2026.
07 / Key questions
At a glance
Release status and benchmark findings, with the evidence kept in scope.
What did Mistral release?
A public API preview announced October 6, 2026. Mistral describes it as a text-and-image mixture-of-experts model with one trillion total and 49 billion active parameters.
How did it score?
Artificial Analysis gave Large 4 Preview 38 Intelligence Index points in its October 7 snapshot. This is not a percentage or task guarantee.
Are the weights available?
Not at the time covered by the report. Mistral scheduled their release for later in October; whether that schedule was met is not established here.
Does the score prove it cannot handle agentic tasks?
No. The index offers comparative evidence, but it does not prove failure on every agentic task. Workload-specific testing is still needed.
What the Benchmark Gap Shows
The results matter to developers choosing a model for multi-step work, where systems may need to plan, use tools, interpret outputs and carry decisions forward. Weak assumptions early in a workflow can affect later steps, and a polished final response does not by itself show that the work was sound. The benchmark provides one comparative signal, but it cannot settle how Large 4 will perform on an individual team’s coding, research or business tasks.
The source report’s author, Thorsten Meyer, said he would not choose the current preview for demanding agentic work or long tasks when stronger alternatives are available. That is a reviewer’s judgment, informed by the benchmark and personal use—not a controlled trial or a finding that the model fails every such task. Mistral’s own claims about agentic coding and specialized professional work still need workload-specific evaluation.
The comparison also does not support saying that every competitor is ahead. Canada’s Cohere Command A+ scored 13 on the same snapshot, below Mistral. The narrower supported point is that Large 4 trails several leading U.S. models and stronger Chinese alternatives on this index. Those differences can shape which systems developers test first, but benchmark scores alone do not establish overall product value, reliability or suitability.
As an affiliate, we earn on qualifying purchases.
A Preview, Not a Weight Release
The timing and release format affect what can be concluded. Mistral announced an API preview on October 6, and the source report assessed the model the following day. The weights were due later in October, meaning developers could not yet independently download and run them at the time of that assessment. Mistral said it was continuing to improve the model, so results from the preview should not be treated as a final evaluation of a later release.
Mistral’s European infrastructure is relevant to the company’s effort to build AI capacity in Europe. But where a model is trained or a developer is based does not show where an API request is processed. The comparison’s developer locations refer to the organizations behind the models, not the locations of particular requests. For now, the practical question is how the preview performs against alternatives on actual workloads—not what its parameter count or regional origin implies.
Meyer also reported encountering hallucinations in his own use of Large 4 and said that this reduced his confidence in assigning it longer tasks. He explicitly framed this as personal experience, not a controlled comparison; it does not establish hallucination rates or show that competing models no longer make unsupported claims. It is evidence of one reviewer’s experience, not a general performance measurement.
“I would not choose it for demanding agentic work or long tasks when stronger models are available.”
— Thorsten Meyer, ThorstenMeyerAI.com
As an affiliate, we earn on qualifying purchases.
What the Preview Cannot Establish
The benchmark is a dated snapshot, and model scores may change as systems and evaluations are updated. The source comparison does not provide identical reasoning settings or compute budgets across all models. It also does not show how Large 4 performs on a particular developer’s tasks, how often it makes unsupported claims under controlled conditions, or how reliably it completes long agentic workflows.
The report’s excerpt does not provide the full cost comparison it begins to discuss, so it is not possible here to establish whether Large 4 offers a price advantage. Nor does it include results from a controlled side-by-side test of hallucinations or sustained task completion. Mistral’s claims about agentic coding and professional tasks remain claims to test against specific workloads. The planned weight release had not occurred at the time covered, and its timing and performance could not yet be confirmed.
As an affiliate, we earn on qualifying purchases.
Weight Release and Workload Tests
The next stated milestone is Mistral’s planned release of Large 4’s model weights later in October. That release would allow a different kind of evaluation from the current API preview, though the source material does not confirm that the schedule was met or specify the eventual license and access terms.
For developers, the immediate next step is to test the available preview on the tasks they expect it to handle, checking accuracy, unsupported outputs and performance across complete workflows. Future benchmark results may change the comparison. Until those results and broader task evaluations are available, the October 7 index should be read as evidence about one benchmark snapshot—not a final verdict on Mistral Large 4.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did Mistral release?
Mistral announced Large 4 in public API preview on October 6, 2026. It is a text-and-image mixture-of-experts model with one trillion total parameters and 49 billion active parameters, according to the company.
How did Mistral Large 4 score?
Artificial Analysis gave Large 4 Preview an Intelligence Index score of 38 in a snapshot dated October 7, 2026. The score is a benchmark measure, not a percentage or a guarantee of performance on a particular task.
Are the model weights available?
Not at the time covered by the source report. Mistral said the weights were scheduled for release later in October; whether that schedule was met is not established here.
Does the benchmark prove Large 4 cannot handle agentic tasks?
No. The index provides comparative evidence, but it does not prove that Large 4 will fail a specific coding, research or multi-step task. The source author’s recommendation against using the preview for demanding long tasks is an individual assessment, not a controlled finding.
What remains to be tested?
Developers still need workload-specific evidence about accuracy, hallucinations, sustained execution and cost. The source material does not provide a controlled comparison for these measures or a complete cost analysis.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
