On October 6, 2026, Mistral AI officially released the public beta preview of Mistral Large 4 — internal codename ML4. With 1.05 trillion total parameters, 5.2 billion active parameters and a 160-million-parameter vision encoder, and supporting a context window of up to 1 million tokens, it is the largest model in Mistral's history. Between launch day and the following day, Mistral updated its officially published active-parameter figure from 4.9 billion to 5.2 billion, as documented in a Hugging Face community post.
Mistral Large 4 deeply reworks the Mixture-of-Experts (MoE) architecture. Traditional MoE routes tokens to a handful of large expert modules, whereas Mistral's official blog describes ML4 as "granular MoE" — the router selects from a larger number of smaller, more specialized units, with the goal of keeping inference-time compute at the level of 5.2 billion active parameters while drawing on a richer store of knowledge from the 1.05 trillion total parameters. This design logic is closely aligned with the direction the DeepSeek and Qwen families have explored in the MoE race over the past two years: use sparse activation to drive down inference costs while maintaining or even exceeding the intelligence ceiling of dense models.
On the published benchmark figures, according to beam.ai and several technology media outlets, ML4 scored 61.7% on DeepSWE v1.1, an agentic coding evaluation aimed at real codebases; in the same run GLM-5.3 scored 61%, DeepSeek-V4-Pro 57%, Qwen 3.8 Max 51% and Reflection AI's Beam 44%. On the finance-focused FinWorkBench evaluation, ML4 matched DeepSeek-V4-Pro at 67%. The evaluation firm Artificial Analysis, whose Intelligence Index aggregates 10 benchmarks, concluded that the ML4 preview scored 38, ranking 64th among 225 models entered and sitting behind the "strongest open-source model" tier. Mistral's official blog explicitly states that these are preliminary preview-period numbers and that the model "is still improving rapidly."
The training infrastructure itself is a strategic statement. Mistral's official blog discloses that ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs, with the compute cluster deployed entirely in Mistral's own European data centers, and that the training data covers more than 160 languages, including all official languages of the European Union. Mistral is using hardware autonomy to vouch for itself with European regulators and enterprise customers — data does not pass through US clouds, the model does not depend on US compute, and inference is not subject to changes in overseas terms of service. Mistral's announcement explicitly states that the version deployed in Europe is "operated independently, without reliance on other digital service providers, in compliance with European law."
Cybersecurity is the flagship scenario Mistral leans into hardest in this release. The official blog states plainly that in highly sensitive vertical industries such as cybersecurity, content moderation by closed-source providers can block legitimate vulnerability research and incident response, whereas ML4's combination of open weights and private deployment lets organizations operate under their own security policies. This is not merely a feature point but a complete value proposition aimed at government and large-enterprise security teams. For this reason, before the weights were formally opened, Mistral invited leading cybersecurity partners and national agencies to take part in red-team testing in advance, under a configuration with "reduced content moderation and expanded cybersecurity capabilities."
For developers and enterprises choosing a platform, several practical constraints in the current window must be clarified. First, as of October 22, no downloadable ML4 weight checkpoint had appeared on Mistral's Hugging Face organization page; the model is accessible only through the Mistral Studio API, with weights scheduled to open at the end of October. Second, the context window differs across platforms: Mistral's official page lists 1 million tokens, while the OpenRouter catalog showed 512K (524,288) on the same day. This means developers must test against the specific service endpoint they actually use before building systems on ML4. Third, Mistral describes the model's visual grounding capability in visual reasoning as surpassing some closed-source frontier models, but this claim currently comes only from the official blog, with no independent third-party benchmark yet corroborating it.
Viewed laterally, the competitive coordinate system ML4 occupies is already unrecognizable from two years ago. Before 2024, open-source models were widely regarded as "cheap alternatives" to closed-source frontier models — a performance discount in exchange for a price advantage. Now, more than one open-source model genuinely approaches or surpasses closed-source flagships on specific tasks: the DeepSeek family exported from China a scale-efficiency story that rattled Silicon Valley, the Qwen family keeps breaking new ground in multilingual capability, and the Llama 4 family is climbing steadily on the back of Meta's compute investment. ML4 adds to this coordinate system a new pole that comes from Europe, emphasizes data sovereignty, and stakes out a clear position in high-compliance industries such as cybersecurity, law and finance. In its coverage, The Next Web framed the event as "Europe challenging China's lead in open-source AI," a narrative frame that reflects a reality: competition among open-source flagships is no longer a binary "open vs. closed" opposition but a multipolar contest in which several geopolitical and technical routes contend in parallel.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接