AI・テクノロジー
Mistral "Large 3" Officially Launched — European LLM Resets the Standard for Multilingual Reasoning
機械翻訳 / Machine-translated
機械翻訳 / Machine-translated
On August 29, 2026, Mistral AI officially released its latest flagship model, "Mistral Large 3." The headline changes are a 128K-token context window and enhanced multilingual reasoning across 32 supported languages. The company claims that benchmark results are now competitive with GPT-5 Turbo and the Claude Sonnet 4 family, suggesting that a practically viable option for guaranteeing data sovereignty has emerged for European enterprises operating under the fully enforced EU AI Act.
Mistral AI announced the general availability of Mistral Large 3 on its official blog. The API is immediately accessible via la Plateforme, and availability through Azure AI Studio and AWS Bedrock began on the same day.
Key specifications:
Pricing is $2.00/1M tokens for input and $6.00/1M tokens for output. This is slightly higher than GPT-5 Turbo's $1.50/$5.00, but an EU in-region data processing guarantee plan is available at a comparable price point.
Reactions on X:
"Mistral Large 3 clearly shows improved instruction-following in Japanese. Tasks that used to require fine-tuning now feel workable out of the box."
"Being able to choose EU in-region processing as a default is a differentiator no other vendor offers. When you factor in GDPR and EU AI Act compliance costs, you really can't afford to leave it off your shortlist."
Since its founding in 2023, Mistral AI has maintained its positioning as a "competitive LLM from Europe." Following releases of Mistral 7B, Mixtral 8x7B, and Mistral Large, its cumulative funding from Series A through C had reached approximately €1.1 billion (roughly ¥180 billion) as of end-2025.
Large 3 draws particular attention in the context of the EU AI Act's "high-risk AI" regulations, which entered full force on August 1, 2026. Whether a vendor can legally guarantee in-Europe data processing has begun to directly affect enterprise procurement criteria, and Mistral is positioned to be its biggest beneficiary.
At the same time, now that Meta's Llama 4.1 family has reached GPT-4o-level visual reasoning, the argument that "open weights are sufficient" has grown stronger. Mistral Large 3's adoption of a licensing structure that sits between fully open and fully closed is widely seen as a strategic decision to make "legal trustworthiness" its primary differentiator.
Mistral AI's official blog reports a score of 8.2/10 on JMT-Bench (a Japanese multi-turn evaluation). However, this figure is based on the company's own measurements, and independent third-party verification has yet to be consolidated. Assessing reliability for real-world Japanese business use will require hands-on validation over the next two to four weeks.
Mistral's "EU Hosted" option is configured so that model weights, inference processing, and logs all remain entirely within EU-based infrastructure. It could represent the fastest path to satisfying the EU AI Act's transparency and log-retention requirements. It is also worth examining for Japanese manufacturers, financial institutions, and medical device companies that deploy AI through their European subsidiaries.
Claiming an Apache 2.0 license while requiring a separate agreement for commercial use is a design approach close to the Meta license used for the Llama 4 family. The growing trend in which "open" and "unrestricted commercial use" are becoming separate concepts raises the cost of contract scrutiny during enterprise adoption. Skipping a prior review by the legal department is not advisable.
Benchmark comparisons across coding, mathematical reasoning, and long-form summarization suggest that Large 3 is roughly on par with GPT-5 Turbo and slightly below the Claude Sonnet 4 family. However, the axes of "guaranteed in-Europe processing" and "self-hosting" are unavailable from other vendors, and more companies are expected to make selections based on criteria beyond raw performance alone.
The most significant shift Mistral Large 3 represents is symbolic: it marks the transition in LLM selection criteria from single-axis performance evaluation to multi-axis evaluation.
Once data sovereignty, cost, regulatory compliance, and performance all function simultaneously as purchasing criteria, the decision flow of "choose the highest-performing model" breaks down. For Japanese companies with European subsidiaries facing cost pressure from GDPR and EU AI Act compliance in particular, this release could serve as a concrete trigger to begin calculating cost-effectiveness.
What deserves attention is that Mistral has continued to reject the binary of "fully closed vs. fully open." This stance may not win wholehearted approval from either camp in the short term, but it represents a consistent strategy to effectively capture the niche of "commercial domains requiring legal trustworthiness."
Real-world validation cases within Japan are not expected to emerge until October 2026 at the earliest. Companies considering adoption are advised to make "reviewing data processing agreements" their first step — before performance evaluation.
The GA of Mistral Large 3 marks the moment European LLMs were upgraded from "an option" to "a genuine competitor." The question is whether you can evaluate performance, cost, and regulatory compliance simultaneously — and the time has come to revisit the blueprint for LLM procurement.
This article was written by an AI writer (AI News) from the Mirai News editorial team.
