Mistral 3 release: new phase of open-source AI competition
This week Mistral AI introduced Mistral 3, a new family of open-source, multimodal, multilingual models released under the Apache 2.0 licence. The range includes compact Ministral variants at 3B, 8B and 14B parameters alongside the flagship Mistral Large 3, a selective mixture-of-experts model with 675B total parameters.
Mistral 3 introduces significant architectural improvements compared to previous generations, resulting in higher inference speed and better memory efficiency. Performance benchmarks highlight notable gains for Mistral 3 across reasoning and multilingual tasks. It places second in the LMSYS Arena’s open-source, non-reasoning category and effectively handles over 40 languages, showcasing its multilingual strengths. Additionally, the 14B reasoning variant achieves an 85% score on the AIME 2025 benchmark, outperforming earlier Mistral models and many open-source peers.
Mistral 3 underscores the accelerating competitiveness of open-source AI amid proprietary advances, positioning the model as a counterpoint to recent closed-model dominance. Google’s Gemini 3 recently topped leaderboards like the Artificial Analysis Intelligence Index, surpassing OpenAI’s GPT family in reasoning, multimodality and agentic tasks through integrated TPU efficiency. Meanwhile, open efforts such as Moonshot AI’s Kimi-K2-Thinking have neared top proprietary scores on benchmarks like Humanity’s Last Exam, signalling open models’ rapid catch-up. In the midst of this, the newly released DeepSeek 3.2 model adds another twist to this landscape. It delivers gold medal–level performance on maths and programming benchmarks such as the IMO, CMO, IOI and ICPC, while undercutting frontier proprietary models on cost by up to a factor of thirty and supporting context windows of up to 128,000 tokens. In more practical terms, DeepSeek's 3.2 release underlines that highly capable, reasoning-focused systems are no longer exclusive to US tech giants, well-funded open players, from China and beyond, are now able to catch up and compete with leading AI labs.
Combined, these developments point to a more plural, globally distributed AI frontier, in which open models increasingly set the tone on capability, cost and accessibility. This shift matters not only for developers and start-ups, but also for policymakers and institutions thinking about sovereignty, competition and broad access to advanced AI systems.