RefGemma 4

Gemma 4

Gemma 4

Google's most intelligent open models, built from Gemini 3 research, optimized for intelligence‑per‑parameter and agentic workflows.

Gemma 4

Overview

Gemma 4 is Google's next generation of open models, derived from Gemini 3 research and technology. The focus is maximum intelligence‑per‑parameter and maximum compute/memory efficiency so the models can:

  • Deliver frontier‑level intelligence on personal computers
  • Run on mobile and IoT devices with tight resource budgets

Capabilities

According to the model page, Gemma 4 is designed for:

  • Agentic workflows
    Build autonomous agents that plan, navigate apps, and complete tasks on your behalf, with native support for function calling.

  • Multimodal reasoning
    Strong audio and visual understanding for rich multimodal applications.

  • Support for 140 languages
    Multilingual experiences that go beyond translation and understand cultural context.

  • Fine tuning
    Improve performance for specific tasks with your preferred frameworks and techniques.

  • Efficient architecture
    Run models on your own hardware for efficient development and deployment.

Performance (Selected Benchmarks)

Gemma 4 variants (e.g., 31B IT Thinking, 26B A4B IT Thinking, E4B/E2B) are compared against Gemma 3 27B IT. Highlights from the page:

  • Arena AI (text) (as of 2026‑04‑02)
    Gemma 4 31B IT Thinking: 1452
    Gemma 3 27B IT: 1365

  • MMMLU (Multilingual QA, no tools)
    Gemma 4 31B: 85.2%
    Gemma 3 27B: 67.6%

  • MMMU Pro (Multimodal reasoning)
    Gemma 4 31B: 76.9%
    Gemma 3 27B: 49.7%

  • AIME 2026 (Mathematics, no tools)
    Gemma 4 31B: 89.2%
    Gemma 3 27B: 20.8%

  • LiveCodeBench v6 (Competitive coding)
    Gemma 4 31B: 80.0%
    Gemma 3 27B: 29.1%

  • GPQA Diamond (Scientific knowledge, no tools)
    Gemma 4 31B: 84.3%
    Gemma 3 27B: 42.4%

  • τ2‑bench (Agentic tool use, retail)
    Gemma 4 31B: 86.4%
    Gemma 3 27B: 6.6%

The model card lists additional benchmarks and breakdowns.

Positioning

Gemma 4 is positioned as:

  • Open: Part of the Gemma family of open models
  • Agent‑ready: Built for tool use and agentic workflows (τ2‑bench numbers are a strong signal)
  • Multimodal: Audio + vision support
  • Edge‑friendly: Designed to run on PCs, mobile, and IoT with high efficiency

The Fireship video (linked above) gives a concise community overview and early impressions of model behavior and tooling support.

Edit this page