openjev/openjev-MLX

OpenJev, MLX 8-bit (Apple silicon, text only)

The language model of OpenJev converted for Apple silicon with mlx-lm (8-bit, group size 64), about 27 GB. Everything about the model, the API and the results is in the main card; this page says what this build is, how it was checked and how to run it.

**

Text only. The converter keeps the language model and drops the vision tower, so this build answers questions about text, JSON and DOM, not screenshots.

Measured, not assumed

The same frozen 10,000 text questions, the same helper code, the same fixed calibration, on a Mac (M5 Max):

test	16-bit model as served on an H100	this MLX build on a Mac
10,000 text questions, 34 public sources	84.03% (8,403)	84.03% (8,403); 72 questions each way, 95% interval −0.24 to +0.22 points

1.5% of individual answers change between the two (150 of 10,000), in both directions. The rule fixed before the run was "no more than 0.5 points below the served model"; it passes. Speed on the M5 Max was about 0.3 s for a short text decision (the run took 80 minutes for 10,000 questions, one at a time).

Run it

uv venv mlx --python 3.12
uv pip install --python mlx/bin/python "mlx==0.32.2" "mlx-lm==0.31.3" \
  "transformers==5.17.0" "openai==3.16.2" "huggingface-hub==1.32.0"

# Download the text-only weights and the two helper files.
mlx/bin/hf download openjev/openjev-MLX --revision ddbabafe9fbf296d767a91ad99717752530a3927 --local-dir openjev-MLX
mlx/bin/hf download openjev/openjev helper/shim.py helper/shim_mlx.py \
  --revision 1c341f65bfe5d50fdb935c71e9739c9e0938d6c4 --local-dir openjev-api

TOKENIZER=openjev-MLX SHIM_MODEL=openjev-MLX \
READOUT_T=0.85 READOUT_NOUL_T=1.829074 READOUT_NOUL_BIAS=0 \
READOUT_TARGETED=1 READOUT_INSTR_STYLE=pyrepr SHIM_STAGGER=1 \
  mlx/bin/python openjev-api/helper/shim_mlx.py \
  --helper openjev-api/helper/shim.py --model openjev-MLX --port 3000

Then call http://localhost:3000/v1/systemone exactly as the main card describes. shim_mlx.py swaps only the helper's model client for an MLX one; prompts, option layout, readout and calibration are the helper's own code.

Licence

Weights: CC BY-NC 4.0 (research and non-commercial use), the same as the main repository. Helper and serving files: Apache 2.0.

OpenJev is an independent project, not affiliated with TypeSafe; Jev is their product.