Yes, for small and local work. Gemma 4 is a family of open models from 2B to 31B parameters under the Apache 2.0 licence, built to run on phones, laptops and single graphics processing units (GPUs). Independent indexes place even the largest size far below the frontier, so use it where privacy, cost or offline use matter more than top quality.
First principles
Memory decides where a model can run. The weights take roughly the parameter count times the bytes per parameter, and at 16 bits each parameter is 2 bytes. That puts the 31B at about 61GB, inside one 80GB GPU. Phones have far less memory, so the sizes aimed at them are the smallest.
A mixture of experts trades memory for speed. The 26B A4B stores 25.2B parameters but uses 3.8B for each token. It needs the memory of a mid-size model and does the work per token of a small one.
Size caps quality. Smaller models score below frontier models on hard reasoning and long tasks, and the indexes below show how wide that gap is for Gemma. Fit and quality pull in opposite directions, and Gemma is built for fit.
Knowledge stops at the training cutoff. A model cannot know what was not in its data. With data up to January 2025, anything later has to come from retrieval.
What it is
Gemma 4 is Google’s open-weight family. The first four sizes were released on 31 March 2026, with a 12B Unified model added on 3 June 2026. All sizes read text and images and return text. The two smallest and the 12B also accept audio.
| Size | Parameters | Context | Audio input |
|---|---|---|---|
| E2B | 2.3B effective | 128K | Yes |
| E4B | 4.5B effective | 128K | Yes |
| 12B Unified | 11.95B | 256K | Yes |
| 26B A4B (mixture of experts) | 25.2B total, 3.8B active | 256K | No |
| 31B Dense | 30.7B | 256K | No |
The model card lists training data up to January 2025 and support for more than 140 languages.
Where it is good
Hardware reach. Google aims E2B and E4B at phones and small devices. The 12B targets laptops with 16GB of video or unified memory. The 31B in 16-bit precision fits on one 80GB H100 GPU. The weights are on Hugging Face, Kaggle and Ollama, and run in llama.cpp, vLLM and Google’s LiteRT-LM.
Licence. Apache 2.0 has no user-count thresholds or naming rules, unlike the custom licences on several rival open models (see our Kimi K3 and Qwen reviews). For a lab or startup building products on small models, this is the cleanest licence in this edition along with Granite, Sarvam and gpt-oss.
Quality for size. Google reports, for the 31B, 85.2% on Massive Multitask Language Understanding (MMLU) Pro, 80.0% on LiveCodeBench v6 and 84.3% on Graduate-Level Google-Proof Question and Answer (GPQA) Diamond. These are vendor-reported. Google also says the 31B ranked third among open models on the Arena text board at launch; that is Google’s citation of an independent board, not a current reading.
Cost. The Gemini application programming interface (API) lists Gemma 4 with free input and output, and third-party hosts sell it. Sarvam, for example, lists Gemma 4 31B at ₹36.60 input and ₹91.50 output per million tokens.
Languages. More than 140 languages makes it a sensible base for fine-tuning in Indian languages, though we found no independent Indic evaluation.
Where it is bad
The gap to the frontier is large. Artificial Analysis gives Gemma 4 31B 19 on its Intelligence Index and the 26B A4B 17. Gemini 3.8 Flash scores 41 and Claude Opus 5.5 58 on the same index. Epoch AI’s Capabilities Index puts the 31B at 142.7, against 157.1 for Gemini 3.8 Flash.
Stale knowledge. With training data ending in January 2025, it knows nothing about most of 2025 and 2026. Pair it with retrieval for anything current.
Audio stops at 12B. If you need speech input and the best quality in the family, you cannot have both.
Agentic work. We found no independent agent benchmark for Gemma 4 on the Vals Index or Arena boards we read. On this evidence, and given the index scores above, it is not the model for long multi-step tool use.
Price and access
Download the weights free. On the Gemini API, input and output are listed as free. Google also offers it in AI Studio and Google Cloud. Your real cost is hardware: a phone for E2B, a 16GB laptop for 12B, an 80GB GPU for 31B at full precision.
Who should use it
On this evidence Gemma 4 is the first family to try for on-device features, offline tools, and small models fine-tuned for one job, such as routing, tagging or extraction. It suits teams that need data to stay on their own machines. For hard reasoning, long coding tasks or current knowledge, call a frontier model instead.
Sources
- Gemma releases | Google AI for Developers, read 2026-09-30
- Gemma 4 model card, read 2026-09-30
- Gemma 4: Byte for byte, the most capable open models, read 2026-09-30
- Introducing Gemma 4 12B, read 2026-09-30
- Gemini Developer API pricing, read 2026-09-30
- Sarvam API pricing, read 2026-09-30
- LLM Leaderboard (Artificial Analysis), read 2026-09-30
- Epoch AI Benchmarking Hub, read 2026-09-30