Gemma 4

Yes. The Gemma 4 model card, the license page linked from the documentation and the April 2 launch post all state that Gemma 4 is released under Apache 2.0, a license that permits commercial use, modification and redistribution. The official Hugging Face repositories carry the apache-2.0 tag and are not gated behind an access request.

This is a change from earlier generations. The Gemma Terms of Use, last modified April 1, 2026, say they apply to the older models listed in their appendix and send Gemma 4 users to the Gemma 4 license instead. Apache 2.0 still has conditions: redistributed copies must include the license text and notices, and modified files must say they were changed.

Google publishes approximate memory needed just to load each size. The figures include 20% overhead for loading but not the context window, which grows with the length of the prompt and reply. The 26B A4B model is a mixture of experts: it activates about 4B parameters per token, but all 26B must be loaded, so it needs almost as much memory as a dense model of that size.

The get-started guide suggests the 26B A4B model as a starting point for many tasks. It lists E2B and E4B for mobile devices and laptops, 12B for laptops, desktops and small servers, and 31B for large servers. The launch post adds that unquantized 16-bit weights of the larger models fit on a single 80 GB NVIDIA H100.

• E2B: 11.4 GB at 16-bit, 5.7 GB at 8-bit, 2.9 GB at 4-bit, 1.1 GB on mobile.
• E4B: 17.9 GB, 8.9 GB and 4.5 GB; 2.5 GB on mobile.
• 12B: 26.7 GB, 13.4 GB and 6.7 GB.
• 26B A4B: 57.7 GB, 28.8 GB and 14.4 GB.
• 31B: 69.9 GB, 34.9 GB and 17.5 GB.

Every size accepts text and images, and all of them can read video as a sequence of frames. Audio input for speech recognition and speech translation is limited to E2B, E4B and the 12B model. The model card caps audio at 30 seconds and video at 60 seconds when processed at one frame per second, so longer recordings must be split.

Context is 128K tokens on E2B and E4B and 256K on 12B, 26B A4B and 31B. For images, a configurable visual token budget from 70 to 1,120 trades detail for speed: the model card recommends higher budgets for OCR and small text, and lower ones for captioning or many video frames.

Google AI Studio offers the 26B A4B and 31B models for trying prompts in a browser, according to the launch post. On the Gemini API pricing page, Gemma 4 input, output and context caching are free of charge on the free tier, while the paid tier is marked not available. The same row says free-tier content is used to improve Google products.

That last point matters more than price. Code, contracts or customer data sent through the free API tier fall under that improvement clause, while the same weights run locally never leave your machine. For production hosting, the launch post points to Vertex AI, Cloud Run and GKE, each billed under its own Google Cloud pricing rather than a Gemma fee.

The documentation links Kaggle and the google/gemma-4 collection on Hugging Face as download sources, and the DeepMind page adds Ollama, LM Studio and Docker. On Hugging Face, each size has a base model, an instruction-tuned -it version and an -assistant draft model used for faster speculative decoding. Search results also show unofficial sites offering Gemma 4 downloads; use only the sources Google links.

The Ollama guide pulls the default model with ollama pull gemma4 and lists the tags gemma4:e2b, gemma4:e4b, gemma4:26b and gemma4:31b. For llama.cpp or LM Studio, the memory page points to the official quantization-aware GGUF checkpoints, which are trained to keep quality close to the full-precision model at 4-bit.

The model card gives one sampling configuration for all use cases: temperature 1.0, top_p 0.95 and top_k 64. Thinking mode is switched on by placing the think control token at the start of the system prompt, and the model then writes its reasoning in a separate channel before the answer. Libraries such as Transformers and llama.cpp apply the chat template for you.

In multi-turn chats, earlier reasoning should not be sent back to the model; only the final answers belong in the history, except for tool-call turns. Put images before the text of a prompt and audio after it. Ignoring these rules does not break the model, but results can differ from the published benchmarks.

1. Record the machine, operating system, Ollama version and model tag.
2. Ask for the function and five pytest cases, with thinking mode left at its default.
3. Save both files and run pytest locally without editing the code.
4. Ask once for a fix if any test fails, then run the tests again.
5. Note the total wall-clock time and whether the network stayed disconnected.

Accept code that passes its own tests and three extra cases you add by hand, including one invalid date. Reject answers that call external packages you did not request, invent date formats or never produce runnable code. No code quality or speed result is reported until the exercise is run.

If you want a finished assistant inside your editor rather than a model to host, compare GitHub Copilot and Cursor. For a hosted general assistant with no hardware to manage, see Claude or ChatGPT.