Mistral Large 4 enters public preview: 1T-parameter open-weight multimodal model, weights due by end of October
Mistral AI has launched a public preview of Mistral Large 4, a 1-trillion-parameter natively multimodal model with 49 billion active parameters. The preview API is live on Mistral Studio, and open weights are promised by the end of October 2026. Pricing and context window have not been disclosed.
Mistral Large 4 at a glance
How Mistral Large 4 ranks now
| Benchmark | Score | Rank | Source |
|---|---|---|---|
| AA Intelligence Index | 26.1 index | 36 of 56 | Artificial Analysis, Nov 2026 ↗ |
| Arena Elo | 1423 Elo | 87 of 146 | LMArena, Nov 2026 ↗ |
| Speed (tok/s) | 57 tok/s | 81 of 105 | Artificial Analysis, Nov 2026 ↗ |
Rank is against every model in our database with a sourced score on that benchmark, each counted once with its best score.
Mistral AI launched a public preview of Mistral Large 4 (ML4) on October 6, 2026. It is a 1-trillion-parameter, natively multimodal model with 49 billion active parameters. The preview API is available now on Mistral Studio, and the company says open weights will follow by the end of October.
Key specifications
- Parameters: 1T total, 49B active per token
- Modality: Natively multimodal (text and images)
- Training hardware: 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters
- Training data: Multilingual, spanning 160+ languages, including every official EU language, according to Mistral
- Context window: Not disclosed
- Pricing: Not yet disclosed
- License and training cutoff: Not disclosed
Mistral says the preview is served from the same European infrastructure used for training. It also says the model was built with the same training, customization and RL environment offered to customers through Mistral Forge.
Benchmark claims
Mistral claims ML4 is competitive with the strongest open models globally and significantly outperforms any open-weight model developed in the US or Europe. These figures come from Mistral's announcement. Several rely on third-party indices (Artificial Analysis, vals.ai) as reported by Mistral. We have not independently verified them.
Coding and agents
- DeepSWE v1.1: 61.7%
- SWE-Atlas-QnA: 59.4%
- Terminal-Bench 4: 28.3%
- Coding Agent Index: 49.8%, which Mistral says is ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max
- AutomationBench (657 business workflows): 59.9%, ahead of Kimi K3, MiMo-V2.6-Pro and DeepSeek V4 Pro
- AA-Briefcase (long-horizon knowledge work): 1,393 Elo
- Surge AI blind human evaluation of coding quality: 3.74 out of 5, ranking second of five models. Claude Opus 5 led at 4.22. Kimi K3 scored 3.59, GLM-5.2 scored 3.40 and GLM-5.3 scored 3.60.
Cybersecurity
- Artificial Analysis Cyber Index: top five globally, per Mistral
- Vulnerability reproduce-and-patch test within that index: 82%, the highest of any model according to Mistral
- Cybench (40 competition-derived challenges): 93%
Mistral says Claude Opus 5.5 and GPT-6 Astra score near zero on the reproduce-and-patch test because they refuse the task. The model's score therefore partly reflects refusal behavior, not only raw capability.
Vision
- Dense 200 visual grounding: 42%, versus 41% for GPT-6 Astra, per Mistral
Mistral also cites document, chart and geospatial use cases. The source text provided to us cuts off before the science and math section.
Release approach
Until the weights ship, Mistral is red-teaming ML4 with cybersecurity leaders, vetted partners and state authorities. These groups access a version with reduced moderation and expanded cyber capabilities. Mistral says further details on architecture, additional benchmarks and post-training methodology will follow. It also says ML4 will serve as the base for a new generation of specialized Mistral models.
The model will be offered in multiple regions, including a European deployment that Mistral says it operates end-to-end under European law.
What this means
A 1T-parameter model with 49B active parameters puts Mistral in the same size class as the largest open-weight releases from Chinese labs. If the weights ship on schedule, it would be the largest open-weight model from a US or European developer. Whether the weights are usable at that size depends on a license that has not been published.
The cyber positioning is the strategic bet. Mistral is selling the absence of provider-side refusals, plus self-hosting, as a feature for defenders. That claim is testable once weights are public, and it will draw scrutiny over misuse risk.
The coding results are mixed. The model places second in a human evaluation but sits behind Claude Opus 5. A 28.3% Terminal-Bench 4 score suggests long-horizon terminal work remains hard. Without a stated context window or pricing, teams cannot yet calculate deployment cost. Those two numbers, plus the license, are the main open questions before the weights drop.
Related Articles
Thinking Machines Lab releases Inkling 2 Mini: 26B-parameter looped MoE with 3B active and 1M-token context
Thinking Machines Lab has released Inkling 2 Mini, a multimodal Mixture-of-Experts model with about 26B physical total parameters and 3B active per token. It repeats its middle layers three times, supports a 1M-token context, and ships with a DSpark drafter. Hugging Face reports it leads 32 of 48 benchmarks against four comparable models.
Tencent releases Hy4 flagship weights: 770B-parameter MoE, 1,048,576-token context, conditional license
Tencent released downloadable weights for its flagship Hy4 on November 20, with a System Card. The multimodal MoE has a 770-billion-parameter language backbone, 49 billion active parameters per text token and a 1,048,576-token context window, under a license with commercial conditions for model-as-a-service operators.
Google releases Gemini 4 Neon at $3/$15 per 1M tokens with a 1M-token context window
Google released Gemini 4 Neon, available today through the API at $3 per million input tokens and $15 per million output tokens. It is the fourth new Gemini 4 model in four weeks, with a context window of about one million tokens.
KAIST MLP Lab releases KORMo-Web-11B: open LLM, VLM and web agent scoring 30.1 on Korean Online-Mind2Web
KAIST MLP Lab has released KORMo-Web-11B, an 11B-parameter bilingual model that combines text, vision and browser-agent capabilities in one set of weights under Apache-2.0. On 612 Online-Mind2Web tasks it scores 29.0 in English and 30.1 in Korean, according to the lab.
Comments
Loading...