jevos-v4

Yes/no decisions on a laptop CPU in 25–110 ms: jev, one binary for Windows, Linux and macOS, now with jevos-v4 at 8-bit weights. 87.0% on 821 new yes/no questions written by another model family (jevos-v3 86.5%), better scores (58.5% on 2,350 held-out score questions, jevos-v3 56.5%; 63% on everyday ratings, 61%) and choices (79.2%, 78.8%). On the 999 hand-written yes/no questions it gets 78.9%, below jevos-v3's 80.8%: if that set is closest to your use, jevos-v3 stays available in its release. Same size and speed as jevos-v3.

jevos-v4 is the weight average of a new training run and jevos-v3. The new run used the same 17-layer model with a wider training mix: jevos-v3's data plus structured records, long written policies and public decision tasks, with uncertain teacher labels filtered out. Across our 24 benchmark sets it averages 67.8%, against 66.2% for jevos-v3. Measured through jev itself (OpenVINO INT8).

Download the archive for your system and the model, and unpack the model into the jev folder:

tar -xzf jev-linux-x64.tar.gz                 # Windows: unzip jev-windows-x64.zip
cd jev
unzip ../jevos-v4-openvino-int8.zip           # creates model/
./jev serve                                   # Windows: jev.exe serve

File	What it is
jev-windows-x64.zip, jev-linux-x64.tar.gz (glibc 2.35+), jev-macos-arm64.tar.gz	the jev binary with OpenVINO's libraries and the licenses
jevos-v4-openvino-int8.zip	the model jev runs, as model/
jevos-v4-q4_k_m.gguf, jevos-v4-q8_0.gguf	the same model as GGUF, for llama.cpp and the tools built on it
jevos-dino.mp4	the recording of jevos playing a Dino-style game
SHA256SUMS.txt	the SHA-256 of every file here

The API, options and numbers are in the README.