Gev E4B
Google's Gemma 4 E4B, retrained to answer questions about a text with
probabilities: which team should handle a ticket, how urgent it is, whether a
rule allows a case. It is the default model of
Gev, which serves it through an HTTP API
compatible with Jev's /v1/systemone.
The model is meant to be run by Gev, which reads the probability of each answer at one position instead of generating text. It is not tested as a chat model.
Files
| File | Use |
|---|---|
model.safetensors, config.json, tokenizer files | the weights, bf16, for building a Mev engine on a Mac (scripts/build-engine.sh gev-e4b) or for your own tools |
gev-e4b-q8_0.gguf | for llama.cpp, 8.0 GB; Gev's Docker image downloads it |
gev-e4b-q4_0.gguf | for llama.cpp on smaller GPUs, 5.2 GB, laid out like Google's QAT Q4_0 GGUF |
Retraining moved the weights off the grid Google's quantization-aware training had put them on, so Q4_0 costs this model more than it costs the original. On the 316 example requests of Gev's playground, the most probable answer through llama.cpp matched the one from the unquantized weights for 99.1 % of the questions with Q8_0 and 93.9 % with Q4_0 (Google's E4B at Q4_0: 96.2 %).
A ready-built engine for Apple's M4 is in hilman2/gev-e4b-mev.
Results
Measured on a Mac mini with an M4 Pro and 48 GB, through Gev's Mev engine.
| Gev E4B | Gemma 4 E4B | Gemma 4 26B-A4B | |
|---|---|---|---|
| Correct decisions in 7,671 test cases | 88.8 % | 83.6 % | 88.2 % |
| Requests per second, 8 at a time | 4.54 | 2.52 |
License
Apache License 2.0, as Gemma 4.