Model releases · 11 August 2026
Google published separate quantisation-aware-trained variants of its two smallest Gemma 4 models, aimed at mobile deployment — a different thing from taking the standard weights and compressing them afterwards.
| Released by | Google DeepMind |
|---|---|
| Published | 20 July 2026 |
| Variants | E2B and E4B, each in -ct and -transformers form |
| Architecture | Gemma4ForConditionalGeneration (gemma4) |
| Context length | 128K on the small models, 256K on medium |
| Modalities | Text and image; audio supported on E2B, E4B and 12B |
| Languages | More than 140 |
| Licence | Apache 2.0 |
| Weight size | 2.67 GB for E2B QAT mobile |
| Hosted at | google on Hugging Face |
Most small models reach a phone through post-training quantisation: the lab publishes full-precision weights and someone compresses them afterwards, accepting whatever quality is lost. Quantisation-aware training puts the compression inside the training loop, so the model learns weights that survive it.
The practical difference is that a QAT four-bit build is generally closer to its own full-precision parent than a post-hoc four-bit build of the same model. Google publishing these as separate repositories, rather than as a note in the main model card, is a signal about how it expects Gemma 4 to be deployed.
Gemma 4 interleaves local sliding-window attention with full global attention, with the final layer always global. Google's framing is that this gives the speed and memory profile of a lightweight model while keeping long-context awareness. To hold memory down at long context, the global layers use unified keys and values and apply Proportional RoPE.
The family is multimodal — text and image input across the line, with audio supported on E2B, E4B and 12B — and covers more than 140 languages with native support for a system role.
Four repositories: E2B and E4B, each in a -ct and a -transformers flavour. The E2B QAT mobile weights are 2.67 GB. All are Apache 2.0, which puts them among the more permissively licensed models available in this size class.
| Runtime | Supported | Notes |
|---|---|---|
| llama.cpp | Yes | The gemma4 architecture has upstream support and community GGUF builds are widely available. |
| transformers | Yes | The -transformers variants are published for exactly this path. |
| MLX | Partial | Community conversions; nothing published by Google. |
| Audio input | Partial | Supported on E2B, E4B and 12B, but runtime support for audio lags behind text and image. |
Training that includes the compression step in the loop, so the model learns weights that survive being reduced to four bits. It generally holds quality better than compressing full-precision weights after training.
Apache 2.0, including the QAT mobile builds.
Text and image across the family, with audio supported on the E2B, E4B and 12B sizes. Runtime support for audio varies by engine.
2.67 GB.
128K tokens on the small models and 256K on the medium ones.
OnDevice LLM is a private AI assistant that runs entirely on your iPhone — no account, no cloud, and nothing you type leaves the device.
Published 11 August 2026