Gemma System One
github.com/zozo123/gemma-to-jev- Category
- Developer Tools
- Rank
- No. 4112Tools index
Previous survey · No. 3767 ·
- Type
- TOOL
- Use case
- Models: Train & Run
- Date
About
Gemma System One turns Google's Gemma 3 4B model into a typed decision function in native Rust using Candle, skipping open-ended text generation entirely: a single forward pass produces logits over the legal answer options, which are softmaxed into a typed answer instead of generated, parsed text. On an Apple M1 Pro it measures 47ms per warm decision (21.2 decisions/sec) and scores 58.9% on JevBench's 231 public decisions.
At a glance
Gemma System One uses Gemma to score a fixed set of possible answers directly. It is worth studying if you need fast, structured decisions and want to measure how sharing inference state affects the answer.
Good fit
Inference researchers comparing latency and decision stability on a fixed task.
Consider something else when
Use a conventional classifier if you need established calibration or a supported production runtime.
Start here
Read the isolation experiment and compare the same choices across the default isolated path and the experimental packed path. Check agreement before interpreting any latency difference.
Cost and requirements
The documented demo uses Rust, Candle and Gemma 3 4B weights. The project reports a roughly 2.3 GB initial weight download and tests its featured path on Apple Metal.
Limits to know
The project says packed inference can change answers. Its published latency comes from that experimental path. A normalized choice score does not by itself establish calibration.
Sources checked 2026-11-05. Source-based assessment.
Why it made the leaderboard
This library turns Gemma 3 4B into a typed decision function via a single forward pass over legal-label logits instead of token generation, giving native Rust apps a 47ms, no-parser way to get structured LLM decisions.
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.