Gemma 4

google's open-weight (open-weight-models) model family — the locally-runnable, Apache-2.0 counterpart to its proprietary Gemini line. A multi-size lineup tuned for on-device / efficient deployment.

Family (2026)

The report also fixes the bottom of the range at 2.3B, and describes the suite as mixing dense and MoE architectures without saying which sizes are which gemma-4-technical-report.

Notable

• Encoder-free multimodality (12B): vision via a single matrix-mul embedding module; audio projected directly into the text token space — no separate modality encoders gemma-4-12b-announcement. It is a one-model choice, not a family-wide one: every other size got improved vision and audio encoders instead gemma-4-technical-report.

• Thinking mode: Gemma 4 generates reasoning traces before answering gemma-4-technical-report — Google putting explicit reasoning in the open track, not only in gemini. Bears on the wiki's open question of whether open weights close on frontier reasoning.

• MoE at the top end (26B-A4B): activates 3.8B of 25.2B params — the sparse-MoE pattern that dominates 2026 open weights (open-weight-models, cf. llm-inference).

• Licensing: Apache-2.0 — among the permissively-licensed open-weight leaders.

• quantization (QAT): official quantization-aware-trained checkpoints for E2B / E4B / 26B-MoE — Q4_0 (4-bit) plus a targeted 2-bit mobile scheme; the E2B text model fits in < 1 GB, claimed higher quality than PTQ gemma-4-qat. Footprint is an explicit lever.

• Deployment: Hugging Face, Kaggle (GGUF + compressed-tensor); runs via llama.cpp, Ollama, LM Studio, MLX, vLLM, SGLang, Transformers.js, LiteRT-LM, Unsloth — broad local/edge support. Open weights → self-host, not a metered API (llm-api-pricing). But self-hosting is not the only route: 14 hosted providers resell Gemma 4 as an API, from Vertex and Cloud Run to Cerebras, Together, Fireworks, Baseten, Modal, Runpod and Cloudflare Workers AI awesome-gemma. Open weights multiplied the hosting market rather than removing it.

• A family of families: 16 first-party domain variants carry the Gemma name — MedGemma, PaliGemma 2, TxGemma, VaultGemma and the rest — several of them built on Gemma 2 or 3 rather than on this generation. See gemma-variants; the name spans more than this page does.

• Derivatives edit the safety layer, not just the precision. Apache-2.0 plus downloadable weights means third parties ship modified checkpoints under the same license: gemma-4-31b-it-scotoma is 31B-it with its refusal direction thinned across layers 7–41 (abliteration). Google's alignment travels with the weights and can be edited out of them.

Place in the market

Gemma 4 is google's entry in the open-weight wave alongside Meta Llama 4, Alibaba Qwen3, deepseek, Moonshot Kimi — competing on local-deploy efficiency and modality rather than raw frontier reasoning. The 12B's encoder-free audio is the family's current differentiator.

That framing is now under pressure from the maker's own account: the gemma-4-technical-report claims Gemma 4 "rivals larger, frontier open models in human-rated tasks" and ships thinking mode. Both claims are Google grading Google, so the efficiency-first reading above stands until an independent bench moves — but the positioning is contested rather than settled. See ../synthesis.md.

Related

gemma-variants · gemma-4-technical-report · awesome-gemma · google · open-weight-models · quantization · gemma-4-qat · gemma-4-12b-announcement · open-source-llms-2026 · llm-benchmarks · deepseek · gemma-4-31b-it-scotoma · abliteration