<body><div>Google releases a new multimodal model Gemma 4 12B, revolutionizing the traditional architecture by eliminating the separate encoder component, achieving effici</div></body>
**Google DeepMind Launches Gemma4: A New Era of Open-Source AI** Google DeepMind has officially unveiled the Gemma4 open-source model series, marking a significant annual update. This release represents a generational leap in performance and a major shift in open-source policy. Google has transitioned the model's license to the globally permissive Apache 2.0, granting developers greater freedom for commercial use and modification. **Complete Model Spectrum: The "Four Pillars"** The Gemma4 series introduces four distinct models for comprehensive scenario coverage: * **31B Dense (Flagship):** Featuring 310 billion activated parameters and 256K context length, it ranks #3 on the Arena AI leaderboard and runs on a single H100 GPU. * **26B A4B MoE (Value Leader):** Utilizing a Mixture-of-Experts architecture (252B total, 38B active parameters), it offers inference speed near a 4B model with superior quality, ranking #6. * **E4B & E2B (Edge-Optimized):** Designed for mobile and embedded devices. Using Per-Layer Embeddings, their effective parameters are compressed to 4.5B and 2.3B respectively, with E2B requiring under 1.5GB memory on some devices. **Performance Breakthroughs: A Generational Leap** Gemma4 demonstrates dramatic improvements over its predecessor, Gemma327B: * **Mathematics:** AIME2026 score surged from 20.8% to 89.2%. * **Coding:** Codeforces ELO rose from 110 to 2150; LiveCodeBench performance increased from 29.1% to 80.0%, establishing it as a top open-source coding assistant. * **Reasoning:** GPQA Diamond score nearly doubled from 42.4% to 84.3%. * **Multilingual:** Natively supports 140+ languages, achieving an 88.4% MMMLU score. **Core Innovations: "Thinking Mode" and Agent Capabilities** Beyond scaling, Gemma4 incorporates advanced interactive features: * **Thinking Mode:** A toggleable feature enabling internal reasoning before output, significantly improving accuracy in multi-step planning tasks. * **Native Agent Support:** Supports function calling and structured JSON output. Google released an open-source Agent Development Kit (ADK) to empower