Google DeepMind Releases Gemma 4 Open Model Family in Four Sizes
Google DeepMind
- Models
- Open source
- Benchmarks
- #gemma 4
- #apache 2.0
- #moe
Google DeepMind announced Gemma 4, an open model family under Apache 2.0, in four sizes: E2B and E4B for edge devices, a 26B MoE, and a 31B Dense model. The publisher claims native function-calling, structured JSON output, system instructions, vision, audio on the edge models, and context windows of up to 256K. Google reports its 31B model ranks #3 and the 26B #6 on Arena.ai's text leaderboard, and cites over 400 million cumulative Gemma downloads.
What Google announced
Google DeepMind introduced Gemma 4, described as its most capable open model family to date, built from the same research as Gemini 3. The release covers four sizes: Effective 2B (E2B), Effective 4B (E4B), a 26B Mixture of Experts model, and a 31B Dense model. The publisher says the edge models handle multimodal input including native audio, while all models process images and video and support long context — 128K for the edge models and up to 256K for the larger ones. Agentic features cited include native function-calling, structured JSON output, and system instructions. Google says the 26B MoE activates 3.8 billion parameters during inference to favor token throughput, while the 31B Dense prioritizes quality and fine-tuning. The weights ship under an Apache 2.0 license, and Google reports day-one support across tools such as Hugging Face, vLLM, llama.cpp, Ollama, and LiteRT-LM. Ranking claims — #3 for the 31B and #6 for the 26B on Arena.ai's text leaderboard — are the publisher's, as is the figure of over 400 million cumulative Gemma downloads.
A scenario for a mobile developer
A developer building an Android field-service app could consider the E4B model to keep an inspection assistant working offline. The publisher's description of native audio input and on-device inference is the premise: a technician could dictate notes and get structured output without a network round trip. A team would need to verify that its target devices meet the compute and memory requirements, since the excerpt gives effective parameter counts but no per-device footprint. Android developers could also evaluate the AICore Developer Preview, which Google describes as a way to prototype agentic flows with forward-compatibility to Gemini Nano 4. This is a conditional plan, not a claim that such an app already works.
How to test it before committing
A team could pick one narrow task, such as converting voice memos into a fixed JSON schema, and run it against a current baseline on the same devices. Useful measures include output validity rate, latency percentiles, and memory and battery use, all compared to that baseline rather than to published benchmarks. Because the announcement emphasizes function-calling and structured output, a trial should log schema violations and malformed tool calls as failure checks. Human review of a sample of outputs remains necessary, especially for any downstream action taken from model output. For the larger models, a team could repeat the test locally on quantized weights and confirm that quality holds before relying on it.
Trade-offs and open questions
The size split implies trade-offs: the 26B MoE is presented as faster and the 31B Dense as higher quality, so choosing between them depends on whether latency or accuracy matters more for a task. Edge models favor efficiency over raw parameter count, which may limit complex reasoning. Prerequisites include hardware that matches each tier's memory footprint; the excerpt mentions unquantized bfloat16 weights fitting on a single 80GB H100, but not consumer requirements. Unknowns include fine-tuning outcomes on domain data and real-world agent reliability, which the publisher asserts rather than demonstrates in the excerpt. A team should avoid adopting Gemma 4 when it needs independently verified benchmarks, documented compliance evidence, or guaranteed performance on hardware it cannot test.
Context and limitations
- Based on a short excerpt of the announcement; the full article and model card were not analyzed.
- Performance and ranking claims come from the publisher and Arena.ai as cited; no independent testing is reported here.
- Fine-tuning outcomes, latency, and hardware fit for specific tasks are not documented in this excerpt.
- Based on bounded source text, not independent testing. Practical suggestions are editorial proposals.