Google's Gemma 4 ships in five sizes, from a 1.1 GB phone model to a 31B workstation LLM

Google has completed the rollout of Gemma 4, its latest family of open models, in five sizes covering hardware from mid-range phones to workstation GPUs. According to a dev.to writeup by Nokka (which notes it was AI-drafted and human-edited) that assembles Google's announcement pages, model cards and technical report, the first four sizes arrived on 31 March 2026 under an Apache 2.0 licence, with the fifth, 12B Unified, following on 3 June. The whole family is built from the same research and technology as Gemini 3.

A five-size family

The lineup stretches across an unusually wide hardware range. E2B, the smallest, needs about 2.9 GB at 4-bit quantisation, with a mobile build at roughly 1.1 GB and a text-only LiteRT-LM version down to 0.84 GB — small enough for mid-range phones. E4B sits at 4.5 GB (2.5 GB on mobile) for laptops and on-device reasoning. 12B Unified takes 6.7 GB for multimodal workstation use, 26B A4B needs 14.4 GB, and the top-end 31B dense model requires 17.5 GB. All five accept text and images, while audio and video are native to E2B, E4B and 12B. Context windows come in two tiers, 128K tokens for the smaller models and 256K for the rest, with support for more than 140 languages. The dev.to post cautions that these memory figures cover model weights only, so KV cache adds overhead as context grows.

The Mixture-of-Experts sweet spot

The variant drawing the most attention from self-hosters is 26B A4B, a Mixture-of-Experts model with roughly 26 billion total parameters but only about 4 billion activated per token. Google states plainly that it runs nearly as fast as a plain 4B model, which the post frames as the core appeal: capability in the 26B class at small-model speed and memory. It targets consumer GPUs such as the RTX 3090 or 4090, plus Macs with enough RAM.

12B Unified drops the encoders

12B Unified arrived later than its siblings because its architecture differs from the rest of the family. Instead of routing images and audio through separate encoders before the language model, it projects raw image pixels and audio waveforms directly into the embedding space through a lightweight linear layer. The result is one model handling text, images and audio in roughly 6.7 GB at 4-bit quantisation — comfortable on hardware many developers already own.

Benchmark jumps, with caveats

Per Google's own published figures, as cited on dev.to, the 31B model scores 85.2 on MMLU Pro against the previous Gemma 3 27B's 67.6, leaps from 20.8 to 89.2 on AIME 2026 maths, and from 29.1 to 80.0 on LiveCodeBench v6. Google also publishes scores for every size, making the trade-offs visible: E2B manages 37.5 on AIME while 26B A4B reaches 88.3, nearly matching the 31B's 89.2 while using well under half the memory at runtime. At launch, 26B A4B ranked sixth among open models on the Arena leaderboard, with 31B in third. The post adds standard caveats: these are Google's measurements on test sets it chose, thinking mode trades quality for longer generation and more output tokens, and leaderboard positions change weekly.

Ecosystem and Android momentum

Gemma had passed 400 million downloads before this release and reached one billion on 20 August 2026, with more than 100,000 community variants, according to figures Google shared. On Android, Gemma 4 entered the AICore Developer Preview in early April, pairing E4B for heavier reasoning with a faster E2B for quick tasks; Google cites a fourfold speedup and roughly 60 percent lower battery use versus the prior generation. The family also serves as the foundation for Gemini Nano 4 on end-of-year devices, and the Android Studio Quail release from 1 September embeds Gemma 4 as a local tool inside the IDE. A Multi-Token Prediction toolkit released in mid-April can reportedly triple inference speed without quality loss.

Why it matters

Gemma 4 signals a shift in how open models compete: not on raw size or headline benchmarks alone, but on intelligence per parameter and per gigabyte. For self-hosters, the practical map is now clear — E4B through 12B Unified for 16 GB machines, 26B A4B at 14.4 GB as the sweet spot for 24 GB-plus setups, and the slower 31B dense model when maximum quality matters. The Apache 2.0 licence with no user or revenue caps removes the friction that gates several competing open families, which matters for startups embedding models in shipped products. And with a genuinely phone-capable model plus an on-device roadmap through Gemini Nano 4, Google is betting that the next open-model battleground is local hardware rather than the cloud — a challenge rivals such as Alibaba's Qwen team will now have to answer.