TL;DR: Google DeepMind released Gemma 4, its most capable open AI model family, built directly from Gemini 3 research. Native function calling, 86.4% on agentic tool-use benchmarks, and Apache 2.0 licensing — finally a genuinely open model for agent builders.
The Big License News
Before anything else: Gemma 4 ships under the Apache 2.0 license. This is a genuine shift. Previous Gemma versions used a custom Google license that the open-source community widely criticized as too restrictive — particularly for commercial deployments. The Apache 2.0 license is permissive, widely used, and compatible with both research and commercial products. Ars Technica called it "much more permissive and widely used by developers." That's accurate.
This matters practically: teams that were hesitant to deploy Gemma in production due to license ambiguity now have a green light. The model is genuinely open in a way previous Gemma releases weren't.
Built for Agentic Workflows
The Gemma 4 announcement puts agentic workflows front and center. The DeepMind page explicitly lists three primary use cases: autonomous agents that plan, navigate apps, and complete tasks on your behalf, with native support for function calling. Other headline capabilities include:
- Native function calling — essential for any agent that needs to interact with external tools, APIs, or databases
- Multimodal reasoning — audio and visual understanding, supporting richer agent perception
- 140-language support — multilingual agent deployments without additional fine-tuning
- Fine-tuning support — run on your own hardware for domain-specific agent tasks
The Benchmarks That Matter for Agents
The jump from Gemma 3 to Gemma 4 on τ2-bench agentic tool use is 6.6% → 86.4% on the 31B model. That's not an incremental improvement. That's a different capability tier.
- τ2-bench (Retail, agentic tool use) — Gemma 4 31B scores 86.4%; even the 2B model hits 29.4%
- AIME 2026 Mathematics — 89.2% on the 31B model
- GPQA Diamond — 84.3%
- LiveCodeBench v6 — 80%
- MMMU Pro — 76.9%
Deploy Anywhere: Mobile, Desktop, IoT
A core Gemma 4 design goal is efficiency at different compute levels. The model family spans a range of sizes, with the 31B model targeting "frontier intelligence on personal computers" and smaller variants optimized for "mobile and IoT devices." NVIDIA has already collaborated on RTX-accelerated support via Ollama and llama.cpp, making local deployment straightforward on developer hardware.
For teams building on-device agents — the kind that run on smartphones, edge devices, or within constrained enterprise environments — Gemma 4 gives you a capable open model that doesn't require calling a remote API. That's a meaningful constraint removed for latency-sensitive or privacy-sensitive agent applications.
MCP, Function Calling, and the Broader Ecosystem
The Gemma 4 announcement mentions Model Context Protocol (MCP) support alongside native function calling. Ars Technica notes that "AI has shifted toward agentic workflow management in the past year, and Gemma 4 is ready for that change with support for native function calling, structured JSON output, and native multimodal function calling."
Hugging Face has already published an extensive blog on Gemma 4 capabilities, with detailed coverage of fine-tuning paths via TRL, transformers.js, and Unsloth Studio. The open ecosystem is clearly already engaged.
What This Means for Agent Builders
Gemma 4 is Google's most credible move yet into the open agent model space. The combination of Apache 2.0 licensing, strong agentic benchmarks, and broad hardware support gives developers a real alternative to closed API models for building autonomous agents.
Whether you're building customer service agents, coding assistants, research tools, or domain-specific autonomous workers, Gemma 4 is worth evaluating — particularly if you need a model you can run on-device, strong function-calling performance without frontier API costs, or multimodal perception in a lightweight deployable package.