What is Gemma 4?
Gemma 4 is Google DeepMind's newest open model family, built from the research behind Gemini 3 and released under Apache 2.0. It comes in five sizes: E2B and E4B for phones and laptops, a 12B model that takes audio without separate encoders, a 26B A4B mixture-of-experts model that uses 3.8B parameters per token, and a 31B dense model. Every size reads text, images and video frames and writes text; E2B, E4B and 12B also take audio. Its knowledge stops at January 2025.
Unlike Gemini, you download it and run it where you like. Google's own API serves only the 31B and 26B A4B, on a free tier that uses your prompts to improve its products.
What changed from Gemma 3?
Mostly the licence, context and reasoning: Gemma 4 moves to Apache 2.0, doubles the context on its larger sizes, adds thinking, tool calls and audio, and scores far higher on Google's coding tests.
- Licence
- Apache 2.0 replaces the Gemma Terms of Use, so there are no extra use restrictions to pass on to your users.
- Context
- Up to 256K tokens on the 12B, 26B A4B and 31B, twice the 128K of Gemma 3.
- Inputs
- Video frames on every size and audio on E2B, E4B and 12B; Gemma 3 read only text and images.
- Reasoning and tools
- Thinking modes and a native system role are new, and function calling is built into the model.
- Scores
- In Google's table the 31B scores 76.9% on the Tau2 agent test, against 16.2% for Gemma 3 27B.
Who is Gemma 4 best for?
Gemma 4 is best for developers who want a free, Apache 2.0 model on their own hardware: phone and browser apps on the small sizes, and local chat, vision and translation work on the 26B A4B and 31B.
Phone, edge and browser apps
The E2B and E4B models are built for on-device use, and E2B at 4-bit loads in about 2.9 GB of memory, or 1.1 GB on mobile.
Local models on one GPU
The 26B A4B mixture-of-experts model uses 3.8B of its parameters per token, and the 31B dense model fits in 17.5 GB at 4-bit.
Documents, images and audio
Every size reads images at a detail level you choose, and E2B, E4B and 12B also transcribe and translate speech.
Translation and non-English work
It was pre-trained on more than 140 languages and supports over 35 out of the box.
What's included with Gemma 4?
Switchable thinking, native tool calls and system prompts, adjustable image detail, audio input on three sizes, draft models for faster output and official 4-bit builds.
- Switchable thinking
- Turn step-by-step reasoning on or off with a token in the system prompt, or with the thinking level on the Gemini API.
- Tool calls and system prompts
- Native function calling, structured JSON output and a proper system role, which is new in Gemma 4.
- Adjustable image detail
- Five visual token budgets, from 70 to 1,120 tokens per image, trade detail for speed; higher budgets suit OCR and small print.
- Audio input
- E2B, E4B and 12B transcribe speech and translate it into text in other languages, in clips of up to 30 seconds.
- Draft models for faster output
- Every size ships with a matching draft model for speculative decoding, which Google says speeds up output with no loss in quality.
- Official 4-bit builds
- Quantization-aware checkpoints in GGUF and other formats keep 4-bit versions close to full-precision quality.
- Runs in common tools
- Supported in Hugging Face Transformers, llama.cpp, Ollama, LM Studio, vLLM, MLX and Unsloth, and on Vertex AI, Cloud Run and GKE.
How much does Gemma 4 cost?
Gemma 4 is free: Google's Gemini API serves it at no charge with no paid tier, and the open weights for all five sizes download free under the Apache 2.0 licence.
- Input, output and context caching free of charge
- Serves the 31B and 26B A4B; no paid tier or tuning
- Prompts and replies are used to improve Google's products
Because the Gemini API has no paid tier for Gemma 4, what you send there is used to improve Google's products; host the weights yourself to keep data in-house. The weights for all five sizes (E2B, E4B, 12B, 26B A4B and 31B) are free on Kaggle, Hugging Face and Ollama, and loading them takes from about 2.9 GB of GPU memory (E2B at 4-bit) to 69.9 GB (31B at 16-bit), plus more for long contexts. Full pricing on ai.google.dev. We checked these prices on .
How do readers rate Gemma 4?
No one has rated Gemma 4 yet. Be the first.
Sign in to rate. One rating per account.
Gemma 4 pros and cons
It is free, quick on one GPU and strong at translation, writing and vision, but tool calling in coding agents is its weak spot.
Pros
- Free to use and to ship: Apache 2.0 allows commercial use and changes, and Google's hosted API charges nothing.
- Users on Hacker News and Reddit say the 26B A4B runs fast on a single 24 GB graphics card or a recent Mac.
- Many users rate it above rival open models for translation, writing and chat, especially in languages other than English.
- Several users rank its image reading, OCR and handwriting recognition above other open models of its size.
Cons
- Tool calling is its weak spot: many users hit failed or stalled tool calls in coding agents such as OpenCode and prefer Qwen models for agent work, though Google has since fixed some chat template bugs.
- Long contexts need a lot of memory on top of the weights, so users on 24 GB cards fit far fewer tokens than the 256K limit.
- Audio clips are capped at 30 seconds and video at about 60 seconds of frames, only three sizes take audio, and none speaks back.
- Google's own API has only a free tier that uses your prompts to improve its products, and it serves just the 31B and 26B A4B.
Is Gemma 4 open source?
Yes. Gemma 4's code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best Gemma 4 alternatives?
Gemma 4 is an open source alternative to LFM 2.5, Nemotron 3 Nano Omni, TranslateGemma and 6 more.
LFM 2.5FreeLiquid AI's open-weight model family built for edge devices, with tool calling.
Nemotron 3 Nano OmniFreeNVIDIA's small multimodal model with a 256K context.
TranslateGemmaFreeGoogle's open translation models in 4B, 12B and 27B for 55 languages.
LFM3FreeLiquid AI's small text, vision, audio and omni models that run and fine-tune on your own devices.
Nemotron 3FreeNVIDIA's open model family for agents, reasoning and long context.
Falcon H1R 7BFreeTII's downloadable hybrid reasoning model for maths, code and logic.
Questions about Gemma 4
What is the difference between Gemma 4 and Gemini?
Gemma 4 is Google's open model you download and run anywhere; Gemini is Google's hosted family. Gemma 4 tops out at a 256K context and a free API tier that uses your prompts, while Gemini models such as 3.1 Flash-Lite take 1M tokens and offer a paid tier that does not.
Who is Gemma 4 for?
Gemma 4 is best for developers who want a free, Apache 2.0 model on their own hardware: phone and browser apps on the small sizes, and local chat, vision and translation work on the 26B A4B and 31B.
Is Gemma 4 free?
Yes. Gemma 4 costs nothing: the weights for all five sizes are free on Kaggle, Hugging Face and Ollama, and the Gemini API serves it free of charge. The catch on the API is that there is no paid tier, so your prompts are used to improve Google's products.
Is Gemma 4 open source?
Yes. Gemma 4's code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best Gemma 4 alternatives?
The closest alternatives on Recaply are LFM 2.5, Nemotron 3 Nano Omni, TranslateGemma, LFM3, Nemotron 3 and Falcon H1R 7B.
Where can I try Gemma 4?
On its official website, ai.google.dev.
Can I use Gemma 4 in commercial products?
Yes. Google releases Gemma 4 under the Apache 2.0 licence, which allows commercial use, changes and redistribution. That is looser than the Gemma Terms of Use covering earlier Gemma models, which add use restrictions you must pass on to your own users.
Which Gemma 4 size should I use?
Pick E2B or E4B for phones, browsers and small laptops, and the 12B when you want audio input with more capacity. Use the 26B A4B for speed on one GPU, since it uses 3.8B parameters per token, and the 31B for the highest scores, at about 17.5 GB of memory at 4-bit.
Can Gemma 4 run offline?
Yes. Once downloaded, Gemma 4 runs without a connection, and Google says E2B and E4B run fully offline on phones, a Raspberry Pi or a Jetson Orin Nano. The 26B A4B and 31B need a GPU or Mac with about 14.4 GB and 17.5 GB of memory at 4-bit, plus room for context.
Our verdict
Gemma 4 is the open model family to try first if you want free, commercially usable AI on your own hardware, from a phone to a single GPU.
Google's own tables put the 31B far ahead of Gemma 3 on coding and reasoning tests, Apache 2.0 removes the old usage terms, and users rate it highly for translation, writing and vision.
- Skip it for
- Agent-heavy coding work, where many users still find Qwen models more reliable at tool calls.
- Check first
- Test tool calls in your own setup with the latest chat template, and budget memory for long contexts, not just for the weights.
Muse Spark
Command A+
Kimi K3
Qwen3.5 Small
Kimi K2.5
Ring-2.6-1T