Gemini 2.0 Flash

אודות Google DeepMind

הוקם בשנת 2010 · Mountain View, USA / London, UK

Google DeepMind was created in April 2023 by merging Google Brain (founded 2011 inside Google) with DeepMind (founded 2010 in London by Demis Hassabis, Shane Legg and Mustafa Suleyman; acquired by Google in 2014). CEO Demis Hassabis leads the combined unit. The lab is responsible for some of the most influential papers in AI: 'Attention Is All You Need' (Vaswani et al., 2017), AlphaGo (2016), AlphaFold (2018-2021), Chinchilla scaling laws (2022) and the Gemini Technical Report (2023). The Gemini product line started in December 2023 with Gemini 1.0 Ultra/Pro/Nano, expanded with Gemini 1.5 (long context, MoE) in early 2024, Gemini 2.0 Flash announced in December 2024 and generally available in February 2025, Gemini 2.0 Pro Experimental and the Gemini 2.5 family (Pro and Flash) in 2025. Google DeepMind also produces Imagen, Veo, Lyria and the AlphaProteo/AlphaFold series in life sciences. Gemini powers AI Overviews in Google Search, Gemini in Google Workspace and on Pixel/Android devices.

ארכיטקטורה

Sparse Mixture-of-Experts Transformer (natively multimodal, fast tier)

Gemini 2.0 Flash is the speed- and cost-optimised tier of the Gemini 2.0 family, announced in December 2024 and generally available in February 2025. It is a natively multimodal Sparse Mixture-of-Experts Transformer that accepts text, image, audio and video inputs and can emit text plus, in the Flash Experimental variants, images and audio. The model was pretrained on Google's TPU v5p infrastructure on a multi-trillion-token mixture of web text, code, books, image-text pairs and YouTube video frames with a knowledge cutoff in August 2024. Post-training combines supervised fine-tuning, RLHF and reinforcement learning against verifiable rewards, with optimisation focused on making the model 2x faster than Gemini 1.5 Pro while exceeding it on most benchmarks. Gemini 2.0 Flash is the first generally available model in the family with built-in tool use including Code Execution, Google Search grounding and the new Live API for low-latency bidirectional streaming of voice and video. Flash also supports a 1M token context window in production, native image output via the Flash Image (Imagen-integrated) endpoint and structured function calling. The model is the default for the Gemini app's chat experience and is widely used in Google Workspace AI features for summarisation, drafting and meeting notes.

פרמטרים

Undisclosed (sparse MoE, smaller than Gemini 2.5 Pro)

הקשר

1,048,576 טוקנים

יכולות

• 1,048,576 token context window in production
• Natively multimodal: text, image, audio, video input
• Low-latency Live API for streaming voice/video conversations
• Built-in tools: Google Search grounding, Code Execution
• Function calling and parallel tool calls
• Roughly 2x faster than Gemini 1.5 Pro at lower cost
• Native image generation via Flash Image variant
• Multilingual coverage across 100+ languages
• Audio output with multi-voice support in Live API
• Strong long-document Q&A and summarisation
• Best for: real-time voice agents, high-throughput multimodal chat, long-document RAG, mobile assistants.

הדרכה ורישיון

Pretrained on Google's multi-trillion-token corpus of web text, code, books, image-text pairs, audio waveforms and video frames. Knowledge cutoff August 2024. Post-training uses supervised fine-tuning, RLHF and RL against verifiable rewards, with a focus on inference efficiency.

רישיון: Proprietary commercial license via Google AI Studio, Vertex AI and the Gemini app. Commercial use permitted under Google Cloud terms.

בדיקות בטיחות: Reviewed under Google DeepMind's Frontier Safety Framework with internal and external red-teaming, including dedicated tests for the Live API audio/video modalities.