What it is
Gemma 4 is Google DeepMind's open-weight model family, released on 2 April 2026. It launched in four sizes: E2B and E4B for phones and small devices, a 26B mixture-of-experts model that uses 3.8 billion parameters per token, and a 31B dense model. A fifth size, the 12B "Unified", followed in late May; it feeds raw image patches and audio straight into the language model without separate encoders.
Every size reads images and video (as frames, up to 60 seconds at one frame per second); E2B, E4B and 12B also take up to 30 seconds of audio. The small models have a 128K-token context and the 12B, 26B and 31B have 256K. All have a configurable thinking mode, native function calling and a system role, and were trained on data in more than 140 languages with a January 2025 cutoff.
This is the first Gemma release under the standard Apache 2.0 licence instead of Google's custom terms. Google also publishes quantisation-aware 4-bit builds of each size, and the models run on Vertex AI, Cloud Run, GKE and on Android through AICore.
Our review
Gemma 4 is the open family to beat at sizes a single GPU or a phone can hold, and the Apache 2.0 licence removes the legal review that held earlier Gemma releases back. On Google's own card the 31B scores 85.2% on MMLU Pro, 84.3% on GPQA Diamond and 80.0% on LiveCodeBench v6, against 67.6%, 42.4% and 29.1% for Gemma 3 27B. At launch Google said the 31B ranked third among open models on the Arena text leaderboard; that ranking dates from April and newer models have shipped since.
The interesting model for most deployments is the 26B A4B. It activates 3.8 billion parameters per token, so it runs at close to small-model speed while scoring 82.6% on MMLU Pro and 88.3% on AIME 2026 without tools. Its quantisation-aware 4-bit GGUF is 14.4 GB, which fits on a 24 GB consumer card with room for context, and OpenRouter lists it at about $0.08 per million input tokens.
The gaps are in long context and agents. Google's own MRCR v2 result at 128K tokens falls from 66.4% on the 31B to 44.1% on the 26B, so the 256K window is usable only on the largest model, and Tau2 tool-use scores of 68 to 77% sit well below frontier agents. Audio is capped at 30 seconds and the 26B and 31B have none at all, so speech features mean the 12B or the phone-sized models.
The release cadence matters as much as the scores. Google added the 12B, assistant-tuned variants and quantisation-aware builds for every size within two months of launch, and those 4-bit files cut the 31B from about 62 GB in BF16 to 17.7 GB with little extra work for the deployer. That range of first-party files, from 5.2 GB for E4B to 17.7 GB for the 31B, is what lets one team standardise a single family across phone, laptop and server.
Where it earns its place
A platform engineer serving an internal assistant on one 24 GB GPU
Replace a paid API for summarising and drafting internal documents
Run the 26B A4B 4-bit GGUF (14.4 GB plus a 1.2 GB vision file) in llama.cpp or vLLM. Hardware is the only cost, and Apache 2.0 means no usage reporting or licence review.
An Android team at a consumer fintech
Classify receipts and transcribe short voice notes on the phone
E4B through AICore reads images and up to 30 seconds of audio offline, so receipts never leave the device. Expect weaker accuracy than the larger models: E4B scores 69.4% on MMLU Pro against 85.2% for the 31B.
A head of support at a 50-agent SaaS company
Draft first replies to a million tickets a month at low cost
At about 1,000 input and 300 output tokens per reply, the 26B A4B on OpenRouter costs roughly $150 a month. Hosted pricing varies by provider, so check the one you pick.
Before you commit
- Setup
- Download from Hugging Face, then load with Transformers, llama.cpp, Ollama or vLLM; the 4-bit GGUF builds run without a conversion step.
- Integrations
- Native function calling and system role, thinking mode with control tokens, Transformers and GGUF builds, Vertex AI, Cloud Run, GKE and Android AICore.
- Data and privacy
- Self-hosted, so data stays where you run it. Apache 2.0 allows commercial use and fine-tuning with no revenue cap or acceptable-use addendum.
- Cost at scale
- Weights are free. The 31B needs about 62 GB in BF16 (one 80 GB card) or 17.7 GB at 4-bit; 26B A4B is 14.4 GB, 12B 7.0 GB and E4B 5.2 GB at 4-bit. OpenRouter lists 31B at $0.20/$0.40 and 26B A4B at $0.08/$0.26 per 1M tokens.
- Limits
- 128K context on E2B and E4B, 256K on larger sizes; audio up to 30 seconds, video up to 60 frames; January 2025 knowledge cutoff.
What's good
- Apache 2.0 licence
- Five sizes from phone (E2B) to one GPU (31B)
- Official 4-bit builds: 31B in 17.7 GB, 26B A4B in 14.4 GB
- Image and video input on every size
What's not
- Long-context recall drops sharply below the 31B
- No audio on the 26B and 31B; 30-second audio cap elsewhere
- Knowledge cutoff of January 2025
- Agentic tool use well behind frontier models
Best for
Developers building on-device, local or single-GPU apps who want a permissive licence and multimodal input.
Alternatives
Gemma 4: common questions
- Is Gemma 4 worth it?
- Gemma 4 is now the default open family below 32B parameters, because its licence finally matches its quality. We rate it 4.4 out of 5.
- How much does Gemma 4 cost?
- Free weights under Apache 2.0. Hosted on Google Cloud and third parties; OpenRouter lists 31B at $0.20/$0.40 and 26B A4B at about $0.08/$0.26 per 1M tokens.
- What are the downsides of Gemma 4?
- Long-context recall drops sharply below the 31B; no audio on the 26B and 31B; 30-second audio cap elsewhere; knowledge cutoff of January 2025; agentic tool use well behind frontier models.
- Who is Gemma 4 best for?
- Developers building on-device, local or single-GPU apps who want a permissive licence and multimodal input.
- How much does Gemma 4 cost for a team or in production?
- Weights are free. The 31B needs about 62 GB in BF16 (one 80 GB card) or 17.7 GB at 4-bit; 26B A4B is 14.4 GB, 12B 7.0 GB and E4B 5.2 GB at 4-bit. OpenRouter lists 31B at $0.20/$0.40 and 26B A4B at $0.08/$0.26 per 1M tokens.
- Is Gemma 4 safe for company data?
- Self-hosted, so data stays where you run it. Apache 2.0 allows commercial use and fine-tuning with no revenue cap or acceptable-use addendum.
- What are the best Gemma 4 alternatives?
- The closest alternatives we have reviewed are gpt-oss-20b, Mistral Small 4 and Phi-4-Reasoning-Vision-15B.
Answers checked Oct 22, 2026.
Make Gemma 4? Show this rating on your site
