Google Gemma 4 Released as Open Source, Rivaling Top Models with 31B Parameters

At 4:03 AM on April 3, Google DeepMind officially launched its new open-source model series, Gemma4. This release marks exactly one year since the previous version. Google not only achieved a "generational leap" in performance but also made a significant move in its open-source commitment: switching the license from its own proprietary agreement to the widely adopted Apache 2.0. This change allows developers to freely use and modify the models for commercial purposes.

Complete Lineup: Four Models for Mobile to Workstation

Gemma4 launches with four distinct models to cover all scenarios:

31B Dense (Flagship): Features 31 billion fully activated parameters and supports an ultra-long 256K context. It ranks third on the Arena AI open-source leaderboard. The unquantized version can run on a single H100 GPU.

26B A4B MoE (Value Leader): Utilizes a Mixture-of-Experts architecture with 25.2 billion total parameters, of which only 3.8 billion are activated per inference. Its reasoning speed is comparable to a 4B model, yet its quality surpasses similar offerings, securing sixth place on the leaderboard.

E4B & E2B (Edge Elite): Optimized for mobile and embedded devices. Using Per-Layer Embeddings technology, effective parameters are compressed to 4.5 billion and 2.3 billion, respectively. The E2B model can reduce memory usage to under 1.5GB on certain devices.

Performance Leap: Code and Math Capabilities See Major Gains

Compared to the previous generation Gemma327B, Gemma4 shows dramatic improvements in core benchmarks:

Math Competitions: Scores on the AIME2026 test surged from 20.8% to 89.2%.

Programming Prowess: Its Codeforces ELO rating increased from 110 to 2150. On LiveCodeBench, performance rose from 29.1% to 80.0%, making it one of the most capable open-source programming assistants available.

Comprehensive Reasoning: Scores on graduate-level science questions (GPQA Diamond) nearly doubled, jumping from 42.4% to 84.3%.

Multilingual Ability: Natively supports over 140 languages, achieving an 88.4% score on MMMLU.

Core Features: Built-in "Thinking Mode" and Agent Capabilities

Gemma4 isn't just about more parameters; it aligns its interaction logic with the flagship Gemini models:

Thinking Mode: Includes an internal reasoning mode that allows the model to process multi-step plans before delivering an answer, significantly improving accuracy on complex tasks.

Native Agent Support: Supports function calling and structured JSON output. Google simultaneously released an open-source Agent Development Kit (ADK), enabling on-device models to act as "intelligent agents."

Deep Multimodal: All versions support image and video input. Even the smaller models include an additional audio encoder for speech recognition and translation.

Industry Perspective: A "Power Reorganization" in Open Source

Over the past year, domestic open-source models (like DeepSeek, Qwen, and GLM) have evolved rapidly, somewhat diminishing Google's influence in the open-source arena. The release of Gemma4 signals Google's return to the forefront through "extreme edge-side engineering" and "more thorough licensing openness."

Conclusion: When Tech Giants Start Talking "Sincerity"