English | 简体中文

Qev-230M-MLX — text-only fast decision checkpoint for Apple Silicon

This repository publishes Qev's smallest, fastest, text-only model: the unchanged LiquidAI/LFM2.5-230M foundation in backbone/, plus a switchable decision LoRA and candidate pointer head trained by the Qev project. It answers a Snake decision in 43 ms at 1.2 GiB peak MLX memory on an Apple M4 and downloads in about 0.5 GB — but it accepts text only: image or video requests are rejected with a clear error. The default multimodal model is twainsk/qev-450m-mlx.

Qev scores supplied answer options through a Jev / TypeSafe-style API (POST /v1/systemone) with real probabilities, not generated JSON text. Decision inference enables the LoRA; native chat (/v1/chat/completions) disables it and uses the unmodified foundation (~170 tokens/s on an M4).

Use the Qev runtime. A generic Transformers pipeline or mlx_lm.generate does not load the pointer head and switchable adapters.

Quickstart on Apple Silicon

brew tap loadchange/qev https://github.com/loadchange/qev
brew install loadchange/qev/qev
qev pull qev-230m        # downloads and verifies this repository (~0.5 GB) into ~/.qev
QEV_MODEL=qev-230m qev decide "Customer asks for a duplicate charge refund" \
  -i "Is a refund requested?" --noul -q
QEV_MODEL=qev-230m qev serve

From a source checkout: uv sync --python 3.12 --extra mlx, then uv run hf download twainsk/qev-230m-mlx --local-dir models/qev-230m-mlx and pass --model models/qev-230m-mlx. Requirements: Apple Silicon, macOS 14+, Python 3.12+. Validated with MLX 0.32.2 and mlx-vlm 0.7.1; no PyTorch is needed.

Contents

• qev_config.json — Qev format version 3, family: lfm2, runtime: mlx, modalities ["text"].
• decision_adapters.safetensors — LoRA (rank 64, alpha 128) over q/k/v/out_proj, in_proj and the MLP w1/w2/w3 of all 14 layers, kept separate from the foundation.
• pointer.safetensors — the 256-dim candidate scorer.
• backbone/ — LiquidAI/LFM2.5-230M at revision 40cb2ad3…fa45, byte-identical to upstream.
• reports/ — evaluation, provenance and MLX-parity summaries; release_manifest.json lists sizes and SHA-256 digests.

Training

Same recipe and data as qev-450m: 17,892 / 1,120 / 1,380 training / calibration / development questions, one A100 40 GB, 4 epochs, batch 8 × accumulation 2 (4,476 updates, ~16 min), lr 1e-4 cosine, frozen foundation, 16.1 M trained parameters, temperature 4.59479341998814 fitted on the calibration split.

Evaluation

Development set (1,380 questions), PyTorch reference; MLX matches within 6 argmax flips (below).

Closed-loop Snake, 20 seeded games: mean 43.4 food, 0 collisions. Paired against the Qwen3.5-0.8B checkpoint on identical questions: general −4.3 pt (McNemar p=0.002), Snake parity (p=1.0).

MLX vs PyTorch parity on all 1,380 development questions: adapter 6 argmax flips (max probability difference 0.059), bf16 6 flips (0.062), net accuracy unchanged. Apple M4 16 GB: Snake decision p50 60 ms (adapter) / 43 ms (bf16), peak MLX memory ≤1.2 GiB, native chat ≈170 tokens/s in bf16.

Limitations

• Text only. Image and video inputs are rejected (qev-230m accepts text only); use qev-450m for pictures.
• Higher option-order sensitivity: reversing option order flips 16.4% of choices and costs 8.1 pt (85.3% → 77.2%), versus 7.2% / −1.0 pt for qev-450m and 3.6% / ±0 for the Qwen3.5 checkpoint. Prefer qev-450m when answer stability across option orderings matters.
• Chinese training coverage is four synthetic rule families; the 50-question Chinese subset is small.
• API compatibility does not imply matching Jev's model quality or confidence values.

License

The backbone and this Derivative Work are distributed under the LFM Open License v1.0 (LICENSE): commercial use by a Legal Entity with annual revenue of USD 10 million or more is not licensed. The decision adapter and pointer head are new files; the foundation weights were not modified (NOTICE).