🔍 Read the full analysis: What’s Holding Mistral Large 4 Back From The AI Frontier? on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Mistral launched Large 4 in public API preview on October 6, 2026, with a trillion total parameters and 49 billion active parameters. Artificial Analysis gives it an Intelligence Index score of 38, below several leading US and Chinese models; a reviewer also reported hallucinations in personal testing. The benchmark is a dated snapshot, and the reviewer’s observations are not a controlled comparison.
Mistral Large 4 launched in public API preview on October 6, but an assessment published the next day argues it is not yet a strong choice for demanding, long-running agentic tasks. The model scored 38 on Artificial Analysis’s Intelligence Index, behind several leading US and Chinese models, while its weights remain unavailable for download pending a planned later-October release.
According to Mistral’s announcement, Large 4 is its largest model to date: a mixture-of-experts system with one trillion total parameters and 49 billion active parameters. It accepts text and images. Mistral said it trained the model on its own infrastructure in Europe and is continuing to improve it. The current release is an API preview; public access to downloadable weights is scheduled for later in October, not available at the time of the assessment.
Artificial Analysis assigned the preview an Intelligence Index score of 38. In the comparison published October 7, that matched OpenAI’s GPT-6 Luna at maximum reasoning effort and came close to DeepSeek V4.1 Flash at maximum effort, which scored 39. Other listed models scored higher: Anthropic’s Claude Opus 5.5 scored 58, Google’s Gemini 4 Argon 53, OpenAI’s GPT-6.1 Sol 52, Z.ai’s GLM-5.3 45 and Moonshot AI’s Kimi K3 44. Cohere’s Command A+ scored 13, a counterexample to any claim that every major competitor outranks Mistral.
The comparison is a dated benchmark snapshot, not a forecast of success on every task. The models were assessed at different named reasoning settings, and the source says those settings do not represent identical compute budgets. The scores are index points, not percentages of intelligence or predicted task success. The assessment author also reports encountering hallucinations in personal use, but describes that as experience rather than a controlled comparison across models.
What’s Holding Mistral Large 4 Back From the AI Frontier?
A trillion-parameter model has entered public API preview, but its benchmark position and early user feedback leave open questions for demanding, long-running agent work.
Where the preview sits
Artificial Analysis scores are index points, not percentages of intelligence or direct predictions of task success.
Long tasks amplify small errors
Agentic workflows involve linked decisions. A fluent final answer does not guarantee that each step was grounded or correct.
Set the right course
A mistaken early assumption can shape the plan and steer later actions away from the goal.
Interpret the evidence
Tool use adds steps where the model must read results, follow constraints and decide what to do next.
Keep people in the loop
For coding, research and professional work, reliability and review effort matter alongside benchmark scores.
What is known so far
Large 4 is a changing preview. The October 7 observations apply to the version and evidence available at that time.
| Area | Reported status | What it establishes |
|---|---|---|
| Access | Public API preview | Developers could evaluate the API version |
| Weights | Planned for later October | Not publicly downloadable at assessment time |
| Inputs | Text and images | Multimodal input support |
| Context | About 512,000 tokens | Capacity does not prove accurate reasoning over all included material |
| Development | European infrastructure; work continues | Origin and capability are separate questions |
| Reviewer experience | Hallucinations reported | Personal observation, not a controlled comparative rate |
What the snapshot cannot settle
Score, cost and anecdotal experience each answer different questions. None alone establishes performance on a team’s specific workflow.
Useful signals
One aggregate index score, the preview’s stated capabilities, and one reviewer’s account of personal use.
Workload-level results
Repeatable comparisons of long tasks, tool use, factual checks, error recovery, cost and human review under disclosed conditions.
What developers should know
The preview offers an early basis for evaluation. Further development, updated benchmarks and the planned weight release may add evidence.
Can I use it now?
It entered public API preview on October 6, 2026. The weights were not yet downloadable; release was planned for later in October.
How did it compare?
It scored 38. Several listed models scored higher; GPT-6 Luna matched it and Command A+ scored lower. Reasoning settings differed.
Does that prove it will fail?
No. The index is an aggregate snapshot, not a direct test of every task or a forecast of success on a particular workflow.
The Gap on Complex Agentic Work
The score matters because agentic tasks involve more than producing a single answer. A model may need to plan, use tools, interpret results and carry decisions across steps. If an early assumption is wrong, later actions can compound the error, even when the final response sounds coherent. For teams considering autonomous coding, research or professional workflows, reliable execution and the amount of human checking required can matter as much as a headline capability.
The assessment’s author says they would begin with higher-scoring alternatives for demanding, long tasks. That is a judgment about this preview, not proof that Large 4 will fail a particular assignment. Artificial Analysis offers an aggregate comparison, while Mistral’s advertised strengths in agentic coding and specialized professional work would need to be tested against specific workloads. A score of 38 alone does not establish how often the model will make unsupported claims or complete a given workflow successfully.
The comparison also puts Mistral’s European development effort in a competitive frame. Training on infrastructure in Europe is relevant to the company’s position as a European AI provider, but origin and capability are separate questions. Developers choosing a model still need to weigh performance, reliability, cost and deployment requirements against their own tasks.
Popular right now
Vitamix VX1 Professional-Grade Blender, 64 oz, Black
Breville Barista Touch Impress Espresso Machine, Noir (BES881NRE)
Philips 3300 Series Fully Automatic Espresso Machine with Milk Frother, Grinder, Hot and Iced Coffee Presets, SilentBrew, 15-Bar Pressure (E
As an affiliate, we earn on qualifying purchases.
Preview Status and Benchmark Comparisons
Mistral’s October 6 announcement described Large 4 as a preview and said work on the model would continue. That status matters: developers are evaluating an API version that may change, while the promised weight release had not yet taken place when the October 7 assessment was written. The results and experience reported here apply to the preview available at that time, not necessarily to later versions.
Artificial Analysis’s table lists developers’ locations, but those locations do not indicate where a particular API request is processed. The benchmark figures also should not be read as a universal ranking of product quality. The assessment notes that reasoning settings differ and that the comparison is limited to an index score. It reports DeepSeek V4.1 Flash as having approximately comparable benchmark intelligence at a much lower measured cost per task, but the supplied material gives no numerical cost figures to support a fuller price comparison.
The author also cites an approximately 512,000-token context capacity reported by Artificial Analysis. A large context window describes how much material a request can include; it does not establish that a model will accurately reason over all of it. In the same way, the reviewer’s reported hallucinations are relevant to their own selection decision but do not establish a measured comparative hallucination rate.
“I would not choose it for demanding agentic work or long tasks when stronger models are available.”
— Thorsten Meyer, author of the October 7 assessment
What the Preview Cannot Yet Establish
It remains unclear how Large 4 will perform after further development or once its weights are released. The source material does not provide controlled, task-specific evaluations comparing it with competitors on long agentic workflows, nor does it establish how often each model produces unsupported output under comparable conditions.
The benchmark results are a snapshot from October 7, 2026, and may change. The listed reasoning settings do not use identical compute budgets, so the table cannot isolate model capability from evaluation conditions. The assessment also refers to a cost advantage for DeepSeek V4.1 Flash without giving the underlying figures here. Actual cost and reliability for a developer will depend on the model version, task, usage pattern and deployment arrangements.
Weight Release and Further Testing
Mistral said the model’s weights were scheduled for release later in October 2026. That release, if it proceeds on schedule, would give developers another way to evaluate and work with Large 4; the source does not confirm a specific date or final release terms. Mistral also said it was continuing to improve the model.
For teams weighing adoption, the next useful evidence would be repeatable tests on their own workloads: long tasks, tool use, constraint-following, factual verification, error recovery and the amount of human review required. Updated independent benchmarks and disclosed evaluation conditions could clarify whether the preview’s current gap narrows. Until then, the published score and one reviewer’s experience inform a cautious assessment but do not settle how the model will perform across all uses.
Key Questions
Is Mistral Large 4 available to use?
It entered public API preview on October 6, 2026. The assessment said its weights were scheduled for release later in October and were not yet publicly downloadable.
How did Mistral Large 4 score against other models?
Artificial Analysis gave the preview an Intelligence Index score of 38 in data available October 7, 2026. Several listed US and Chinese models scored higher, while OpenAI’s GPT-6 Luna matched it and Cohere’s Command A+ scored lower. The comparison uses different reasoning settings.
Does the score prove Large 4 will fail at agentic tasks?
No. The index is an aggregate benchmark, not a direct measure of success on every workflow. The assessment’s author considers stronger alternatives a better starting point for demanding long tasks, but task-specific performance remains to be tested.
What did the reviewer report about hallucinations?
The reviewer said they encountered hallucinations while using the preview. They described this as personal experience, not a controlled study comparing hallucination rates across models.
Source: ThorstenMeyerAI.com
