Maximilian228/Geode-Gemma-4-E2B

Geode · Gemma 4 E2B for phones

A model pack for Geode, an Android app that runs language models entirely on the phone: no cloud, no account. Gemma 4 E2B is Google's small multimodal model: it reads images, watches video, hears voice and sound, and calls the agent's tools. Geode runs it with its own engine (not LiteRT-LM), from Google's quantization-aware-trained (QAT) 4-bit weights.

Packs

Folder	Weights	Size	Quality (KL vs BF16 / top-1)	For
Q4/	Google's QAT Q4_0 weights, 4-bit with a searched scale per 16 values; vision and audio encoders 8-bit; the MTP drafter 8-bit	3.2 GB (2.94 GiB): 1.3 GB in memory, the per-layer embeddings read from flash	0.213 / 85.1%	any 64-bit phone with 6 GB of RAM or more

The scale search lowers the error below Google's own Q4_0 rounding of the same QAT weights (KL 0.244); the same search on the original, not QAT-trained weights gives 0.296.

What is in the pack: the language model (dense.bin), its per-layer embeddings (ple.bin, read from flash as they are needed), the vision encoder (vision.bin: photos, and video frames at 70 tokens each), the audio encoder (audio.bin: up to 30 s of sound, one token per 40 ms), the multi-token-prediction drafter (drafter.bin, from Google's QAT assistant model: the same answer, every draft checked) and the tokenizer.

Speed on a Snapdragon 720G phone (Redmi Note 9 Pro, the processor only): answers ~7-7.5 tokens/s, up to ~13 with the drafter on code (1.3-1.7x on English text); a prompt ~17 tokens/s; a photo at 280 tokens ~18 s to encode.

On a Snapdragon 8 Elite phone (RedMagic 10 Pro), the default run mode (NPU + GPU): a 1,300-token prompt reads at 1,440 tokens/s on the NPU (Hexagon: whole layers, attention included, on its matrix units), and the answer is written on the Adreno graphics chip at ~42-55 tokens/s with the drafter. A photo is encoded in ~0.9 s and 10 s of sound in ~0.3 s, on the NPU too. Where the NPU is not available, an Adreno 7xx or newer graphics chip takes the prompt instead (380 tokens/s); other graphics chips (Mali, older Adreno) are not used. In Geode's Model settings you can choose which units run it.

The packs are Geode's own format (geode.json describes the files). They are not GGUF and do not load in other runtimes.

Use

1. Install Geode (Google Play beta: geode.tifu.xyz).
2. Models → pick Gemma 4 E2B in Geode's catalog (or Download from Hugging Face → Maximilian228/Geode-Gemma-4-E2B).
3. In a chat: attach a photo or a video, hold the microphone button to say what you want, or attach a sound file.

Voice messages, sound and video are processed on the phone and kept with their chat there; nothing is sent anywhere.

Requirements: Android 11+, a 64-bit ARM processor with the dot-product instructions (ARMv8.2), and free storage for the pack.

License

Apache-2.0, as the base model: this pack is a quantized derivative of google/gemma-4-E2B-it-qat-q4_0-unquantized and its QAT assistant by Google (see LICENSE and the Gemma 4 license). Answers come from an AI model and can be wrong.