# Gemini 3 Gemini 3 is the third-generation [[Gemini]] model family from [[Google DeepMind]]. It succeeds Gemini 2.5 as Google's frontier multimodal [[Large Language Models (LLMs)|LLM]] line and serves as the umbrella for the Gemini 3.x minor releases. ## Variants Minor versions and specialty variants already in the vault: - **[[Gemini 3.1 Flash Live]]** — highest-quality real-time audio/voice model - **[[Gemini 3.1 Flash TTS]]** — controllable, expressive text-to-speech with inline audio tags - **[[Gemini 3.5 Flash]]** — fastest agentic/coding model in the family; beats Gemini 3.1 Pro on coding benchmarks, but the Flash tier is no longer the cheap tier - **[[Gemini 3.5 Pro]]** — reasoning-tier model of the 3.5 generation; announced at I/O 2026, due around June 2026 Other expected family members (Pro, Flash, Ultra, Nano) follow Google's usual tiering convention; check https://ai.google.dev/gemini-api/docs/models for the current list. ## Positioning - Natively multimodal across text, image, audio, video, and code - Very long context windows (inherited from the 2.5 generation, pushed further in 3.x) - Headline competitor to [[Claude Opus 4.7]], [[GPT-5.4]], and the open-weight frontier ([[Kimi K2.6]], [[Qwen]]) - Strongest surfaces: [[Google AI Studio]], Gemini App, [[Vertex AI]], integrations into Google Workspace ## Benchmarks (as of April 2026) Figures reported in competitor launches (Moonshot K2.6 comparison tables; treat directionally): - SWE-Bench Pro: ~54.2% - SWE-Bench Multilingual: ~76.9% - Terminal-Bench 2.0: behind Kimi K2.6 (66.7%) and GPT-5.4 (65.4%) - BrowseComp: **~85.9%** (leading the cited field, including Kimi K2.6 at 83.2%) Gemini 3's strongest axis in third-party tables is agentic web browsing / BrowseComp-style tasks; coding is competitive but not dominant. ## Why it matters - Keeps Google in the three-to-four-way frontier race (Anthropic, OpenAI, Google DeepMind, + the Chinese open-weight frontier) - Tight integration with Google's product surfaces (Search, Workspace, Android, Chrome) is the real distribution moat; model quality is only one axis - Very long context remains a durable differentiator for document-heavy workloads ## Caveats - Google has not published complete model cards at the 3.x minor-version granularity for every variant - Comparison numbers cited here come from competing labs' launch posts, not Google's own benchmarks; independent replication preferred - Gemini tiering (Pro / Flash / Ultra / Nano) changes faster than the model version numbers; check the live docs before choosing a variant ## References - Gemini model list: https://ai.google.dev/gemini-api/docs/models - Google AI Studio: https://aistudio.google.com - Google DeepMind: https://deepmind.google ## Related - [[Gemini]] - [[Google DeepMind]] - [[Gemini 3.1 Flash Live]] - [[Gemini 3.1 Flash TTS]] - [[Gemini 3.5 Flash]] - [[Gemini 3.5 Pro]] - [[Gemini Omni]] - [[Gemini CLI]] - [[Gemini Mobile App]] - [[Google AI Studio]] - [[Vertex AI]] - [[Large Language Models (LLMs)]] - [[AI Frontier Model]] - [[AI Foundation Models]] - [[Claude Opus 4.7]] - [[GPT-5.4]] - [[Kimi K2.6]]