AI・テクノロジー
Mistral Unveils "Mistral 3 Large" — A Turning Point as European LLM Matches U.S. Rivals on Reasoning Benchmarks
機械翻訳 / Machine-translated
On September 14, 2026, French AI company Mistral AI officially released its flagship model "Mistral 3 Large." With an MMLU score of 92.4 — 0.4 points above GPT-4o (92.0) — the model's weights were made available immediately under the Apache 2.0 license. In the sense that "a European open-weight model has reached a level practically indistinguishable from closed U.S. models," this marks a turning point that changes the criteria by which organizations select LLMs.
At 10:00 PM Japan Standard Time on the 14th, Mistral AI published "Mistral 3 Large" on its official blog and on Hugging Face. The model has approximately 123 billion parameters and a context length of 128,000 tokens. Key benchmark results are as follows:
Verification posts from engineers have been appearing in rapid succession on X (formerly Twitter).
"Ran Mistral 3 Large locally. It passes difficult HumanEval problems. Nearly indistinguishable from GPT-4o. And it's Apache 2.0. The fact that this came out of Europe is no small thing." (Engineering account, 14,000 likes)
Mistral AI was founded in Paris in 2023. Its "Mistral 7B," released in September of that year, reset the performance bar for open-weight models, and the 2024 "Mistral Large" series marked the first time a European-originated LLM broke into the enterprise market.
In 2025, as the EU AI Act began its phased enforcement, the company strengthened its sales push to European enterprises by leveraging GDPR compliance and EU data boundary support. In a recent interview, the number of paying enterprise customers was cited as "over 1,200" as of March 2026.
This release marks the first flagship update in 12 months. As an open-weight model, it is only the second instance — after Llama 4 Maverick — to claim "GPT-4o equivalence."
While the Meta Llama series imposes commercial use restrictions on deployments with more than 700 million monthly active users, Mistral 3 Large adopts the Apache 2.0 license. Commercial integration is permitted regardless of scale, significantly reducing the legal review burden for organizations.
The release notes explicitly state that, alongside the model release, Mistral AI simultaneously published technical documentation addressing the high-risk requirements under Annex III of the EU AI Act. The model card discloses training data sources and exclusion lists, proactively establishing eligibility for procurement by European public institutions.
An HumanEval score of 87.6% matches the level of the closed models currently used as the foundation for mainstream AI coding agents. For industries such as finance and healthcare, where the use of external APIs is constrained, local or private cloud deployment becomes a genuinely viable option.
A 0.4-point lead over GPT-4o on MMLU falls within the margin of statistical error. Rather than rushing to frame this as "Europe surpassing the U.S.," the more significant fact is that a European model has, for the first time, reached a level that is practically indistinguishable in real-world use.
What deserves attention more than the benchmark scores is the simultaneous achievement of the three-part combination: Apache 2.0 + EU AI Act compliance + EU data boundary. In regulated industries across Europe and Japan — finance, healthcare, and the public sector — this package is increasingly becoming the de facto condition for procurement decisions. It has been reported that at least five domestic financial institutions have announced their intention to consider adopting domestic or European LLMs since the start of 2026, and Mistral 3 Large arrived at exactly the right moment.
The API price is $2.40 per million input tokens (a 44% reduction compared to Mistral 2 Large). For use cases where local deployment is not necessary and the API suffices, there are now more options when it comes to cost structure.
A European LLM has, for the first time, achieved parity with its peers on benchmark scores. The next question is how well models and agents fine-tuned on Mistral 3 Large will perform in actual business operations. Mistral AI is reportedly planning to offer an enterprise fine-tuning API as early as the fourth quarter, meaning real-world implementation cases are not expected to emerge until late 2026. The "choose between the U.S. or Europe" dynamic in LLM selection is quietly beginning to unravel.
This article was written by the AI writer (AI News) of the Mirai News editorial team.
