Gemma 4 E4B
Description
Gemma 4 E4B is a small open language model from Google DeepMind, released on 2026-03-02. It belongs to the Gemma 4 family, whose smallest models are made, according to the authors, for efficient local work on laptops and even high-end phones. The letter E means "effective": the model works like a 4.5 billion parameter one, while the whole file holds about 8 billion because of extra lookup tables. This card links the instruction-tuned version, which is the one made for chat.
According to the model card, it understands text, images, video and audio, and answers with text. It has a thinking mode that can be switched on or off, built-in support for calling tools, and a context window of 128,000 tokens. The authors say it is pre-trained on more than 140 languages and works out of the box in more than 35. In the authors' own table it is clearly weaker than the larger Gemma 4 models, for example on programming and hard reasoning tests, so it suits everyday questions, summaries, simple code and working with documents and pictures rather than hard tasks.
The weights are public without any acceptance step, and the license is Apache-2.0, which allows commercial use; the license text we read has no territorial restrictions. The model works with Transformers and other common tools, and it is also widely repackaged for desktop programs that run models locally. The card does not give a video memory requirement.
Date of check: 2026-10-21.
Pros and cons
Pros
- Small: works like a 4.5 billion parameter model, made for laptops
- Apache-2.0 license, weights open without an acceptance step
- Understands text, images, video and audio
- Thinking mode that can be switched on or off, tool calling
- Pre-trained on more than 140 languages
Cons
- Clearly weaker than the larger Gemma 4 models on programming and hard reasoning
- Context of 128,000 tokens, smaller than the bigger models of the family
- The card does not state hardware requirements
- Benchmark figures come from the authors, not from independent tests
Key facts
- License
- Apache-2.0
- Commercial use
- Allowed
- Developer
- Google DeepMind
- Interface languages
- English
- Checked on
- Oct 21, 2026
Editorial score
News about Gemma 4 E4B
-
Google DeepMind's Gemma 4 12B enables multimodal AI on laptops with only 16GB RAM, natively handling text, images, and audio while approaching the performance of 26B models. Available on Hugging Face under Apache 2.0. #Gemma4 #MultimodalAI #LocalAI #OpenSource #DeepMind -
Google DeepMind released DiffusionGemma — a text diffusion model built by converting Gemma 4 instead of training from scratch, using <10% of original compute. It refines 256-token blocks in parallel, hitting ~1,500 tokens/sec on H100. Two-stage training (reconstruction + SD·RL merging reinforcement learning with sampler distillation) boosts reasoning quality by ~10 points while quadrupling tokens per step. Bidirectional reasoning enables self-correction mid-generation — solving 85% of Sudoku puzzles and fixing math errors before output finalizes. Tradeoffs: lower peak quality than autoregressive Gemma 4, repetition loops from aggressive step reduction, multimodal formatting issues, and throughput parity with standard LLMs at ~32 concurrent users. Released under Apache 2.0 on Hugging Face as experimental foundation for efficient diffusion research. Already powering multilingual speech recognition and interactive radiology reports. #DiffusionGemma #TextDiffusion #GoogleDeepMind #Gemma4 #AIResearch -
Google releases Gemma 4, its most capable open model family yet under Apache 2.0 license. Four sizes from 2B to 31B parameters run on devices from smartphones to workstations. Built on Gemini 3 tech, they excel at reasoning, math, coding, and agentic workflows with native function calling and JSON output. The 31B model ranks #3 and 26B MoE #6 on Arena AI leaderboard, outperforming models 20x larger. Available now on Hugging Face, Kaggle, Ollama, and Google platforms with broad framework/hardware support. #Gemma4 #GoogleAI #OpenSourceLLM #Apache2 #AIModels