Google DeepMind releases Gemma 4: 4 Apache 2.0 multimodal models from edge to 31B dense

• Google DeepMind released Gemma 4 as four Apache 2.0 open models, E2B, E4B, 26B A4B MoE, and 31B dense, spanning phone-class deployment up to workstation and single-accelerator inference launch thread model lineup.
• The headline features are agentic ones: native tool use, configurable thinking, multimodal input, and up to 256K context on the larger models agentic capabilities SGLang summary.
• Google shipped it straight into the serving stack on day 0, including AI Studio, Hugging Face, Kaggle, Ollama, SGLang, vLLM, and Baseten availability Ollama support vLLM quick start.
• Early leaderboard posts put Gemma 4 31B near much larger open models, with Arena ranking it #3 among open models and Artificial Analysis reporting 85.7% on GPQA Diamond for the reasoning variant Arena ranking GPQA result.
• The more interesting part is packaging, not just the benchmark score: Google also pushed Gemma 4 into LiteRT, AI Edge Gallery, browser demos, and Android Studio, which makes the release feel like an open local-agent stack instead of a single checkpoint drop mobile demo ecosystem extras.

You can read the official launch post, skim Hugging Face's unusually detailed architecture and deployment writeup, and check the vLLM launch post for the hardware matrix. The LiteRT overview already lists Gemma-4-E2B as a featured on-device model, and the Hacker News thread immediately turned into a local deployment discussion.

The family splits cleanly into two edge models and two larger reasoning models. Google positions E2B and E4B for mobile and real-time multimodal use, while 26B A4B and 31B target heavier local reasoning and coding workloads Google's size chart.

Hugging Face's release post fills in the table Google only sketches in the thread:

• E2B: 2.3B effective parameters, 5.1B with embeddings, 128K context
• E4B: 4.5B effective parameters, 8B with embeddings, 128K context
• 26B A4B: 26B total parameters, 4B active MoE, 256K context
• 31B: dense 31B model, 256K context

That effective-parameter framing is the trick behind the small models. The Hugging Face post says Gemma 4 keeps the per-layer embeddings design from Gemma-3n, so the edge models carry larger total parameter counts than their active runtime footprint suggests.

Google's pitch is unusually explicit here: Gemma 4 is meant for agents that plan, navigate apps, search databases, and trigger APIs tool use and context. LMSYS's day-0 SGLang note adds the serving details engineers actually care about, including dedicated reasoning-parser and tool-call-parser support for all four checkpoints SGLang summary.

The capability bundle looks like this:

• Native function and tool calling
• Thinking mode for step-by-step reasoning
• Multimodal input across text and images for all models
• Audio support on the small E2B and E4B variants
• Up to 256K context on 26B A4B and 31B
• Support for more than 140 languages, according to the official materials

The official launch post and Hugging Face breakdown also describe a hybrid attention stack: alternating sliding-window and full-context layers, proportional RoPE on the global layers for long context, and shared KV cache in later layers to cut long-context inference cost.

This was a broad release, fast. Google put Gemma 4 into AI Studio and linked weight downloads from Hugging Face, Kaggle, and Ollama on announcement day official availability. Ollama required version 0.20+ and exposed all four variants with simple ollama run tags on day 0 Ollama day-0 support.

The rest of the serving surface landed quickly too:

• SGLang added day-0 support with Gemma 4 reasoning and tool-call parsers SGLang summary
• vLLM published a launch post and Docker quick start, with support across NVIDIA GPUs, Google TPUs, AMD GPUs, and Intel XPUs vLLM quick start
• Baseten added the full family to its model library and called out OCR, function calling, and 256K context in its launch note Baseten availability

That breadth matters more than another leaderboard screenshot. Open models usually arrive as weights first and production ergonomics later. Gemma 4 showed up with the inference ecosystem already waiting for it.

The early benchmark story is mostly about parameter efficiency. Arena posted Gemma 4 31B at #3 among open models and the 26B A4B variant at #6 Arena ranking. A follow-up Arena post framed the 31B model as 24 times smaller than GLM-5 and 34 times smaller than Kimi-K2.5 by total parameter count while landing in the same neighborhood parameter comparison.

Artificial Analysis put more numbers on that picture. It reported Gemma 4 31B Reasoning at 85.7% on GPQA Diamond, just behind Qwen3.5 27B Reasoning at 85.8%, and noted that Gemma used about 1.2M output tokens versus roughly 1.5M and 1.6M for two nearby Qwen reasoning models Artificial Analysis GPQA token usage.

The 26B A4B result is less dramatic, 79.2% on GPQA Diamond in that same post, but it is the more revealing model architecturally. According to the Hugging Face analysis, it activates only 4B parameters inside a 26B MoE, which is exactly the kind of trade that makes local reasoning models more deployable than their headline size suggests.

Gemma 4 did not stop at server inference. Hugging Face's CEO pointed to a WebGPU demo running fully in-browser through transformers.js browser demo, and Google's LiteRT overview lists Gemma-4-E2B as a featured cross-platform edge model for Android, iOS, web, desktop, and even Raspberry Pi.

Phil Schmid's roundup adds the last piece: Gemma 4 support in Android Studio for local agentic coding, plus hooks into Vertex Model Garden, Vertex fine-tuning, ADK, Cloud Run, GKE, and TPU-backed vLLM paths ecosystem extras. That is a bigger release footprint than the benchmark charts imply.

If there is a memorable detail from launch day, it is that Gemma 4 appeared simultaneously as a model family, a serving target, and an on-device runtime story.