Gemini 3 Pro Model Card [pdf] (storage.googleapis.com)

Google DeepMind published the Gemini 3 Pro model card (Nov 2025), detailing their most advanced natively multimodal, reasoning-focused foundation model. Gemini 3 Pro is a sparse mixture-of-experts (MoE) transformer with a massive 1 million token context window for inputs (text, images, audio, video) and a 64K token output. It was trained on a broad mix of public, licensed, crawled, synthetic and Google product user data, instruction-tuned with RL and human preference signals, and developed on TPU infrastructure using JAX/ML Pathways. The MoE architecture lets the model scale capacity while keeping per-token compute manageable, enabling capabilities like agentic tool use, multi-step reasoning, theorem proving, long-context codebase understanding and complex multimodal tasks. The card reports material gains over Gemini 2.5 Pro across reasoning, multimodal benchmarks and long-context tests and lists distribution via Gemini App, Cloud/Vertex AI, API and Studio. Google also lays out safety work: automated and human evaluations, red teaming, and adherence to its Frontier Safety Framework and GenAI policies. Automated safety metrics are mixed — improvements in tone, multilingual and image-to-text safety but a drop on one automated text-to-text safety metric that Google attributes largely to evaluation changes and false positives. Known limitations include hallucinations, occasional timeouts and a knowledge cutoff of January 2025; prohibited and high-risk uses remain restricted under Google’s policies.