Neural ExpeditionSubscribe
TEST RECORD · ← all field reports

FIELD REPORTS/IMAGE/ISSUE #73

Ask one model about your images and entire documents

Frontier-model releases usually come with a catch: the weights are open but the hardware bill is not. Gemma 4's flagship is a 31B dense model that posts near-frontier reasoning numbers while still targeting a single workstation GPU, and the same family scales down to models that run on a laptop or phone. If you have been waiting for an open multimodal model you can actually deploy, this is the release to test.

MODELGemma 4 31B
PUBLISHEDJuly 13, 2026
READ TIME3 min
TESTED BYNeural Expedition
CATEGORYIMAGE

Field notes

01What it does

Gemma 4 is Google DeepMind's new open-weights family, and 31B is its flagship. It is multimodal: text and images go in, with variable aspect ratio and resolution support, and it reasons with configurable thinking modes, so you decide how much deliberation each request gets. The context window is 256K tokens, and it covers more than 140 languages.

The benchmark jump over the previous generation is unusually large. It scores 85.2% on MMLU Pro, 89.2% on AIME 2026 without tools, and 80% on LiveCodeBench v6, versus 67.6%, 20.8%, and 29.1% for Gemma 3 27B. A hybrid attention design that interleaves sliding-window layers with global ones keeps memory low enough for those numbers to fit on a workstation card.

The release also brings native system-prompt support and function calling, which the Gemma line previously lacked. Those are the two features that matter most if you want to drop it into an agent stack. And if 31B is more than you need, the same family ships an MoE variant (26B total with 3.8B active, so it runs at near-4B speed) and phone-class E2B/E4B models with native audio input.

02How to try it

The weights are on Hugging Face under Apache 2.0 in pre-trained and instruction-tuned variants, with day-one transformers support. Load `google/gemma-4-31B-it` and it works like any chat model, images included. Community demo Spaces are already appearing on the model page if you want a browser taste before committing GPU time.

For lighter hardware, start with the 26B A4B MoE (it runs at the speed of a 4B model) or the E4B, then move up. The whole family shares the same interface, so a pipeline built on the small one ports to the flagship unchanged.

03Caveat

31B dense still wants a serious GPU, so plan on a quantized build for 24 GB cards or step down to the MoE variant. Audio input is only native on the small E2B/E4B models, not the flagship. And it is a brand-new release: quantized community builds and fine-tunes are still landing, so expect the deployment recipes to improve over the coming weeks.

04What you can do with it

  • Ask questions about screenshots, documents, and photos with a fully local model.
  • Build agents on native function calling and a real system role, no prompt hacks.
  • Push 256K tokens of context: whole codebases, long transcripts, document sets.
  • Serve 140+ languages from one open checkpoint.
  • Prototype on the laptop-class E4B, then swap in the 31B without changing code.

Try the demo

View model page

Read this issue on neuralexpedition.com →