English | 简体中文

Qev-450M-MLX — LFM2.5-VL decision checkpoint for Apple Silicon

This repository publishes Qev v0.4.0's default model: the unchanged LiquidAI/LFM2.5-VL-450M foundation in backbone/, a switchable decision LoRA (decision_adapters.safetensors) and a candidate pointer head (pointer.safetensors) trained by the Qev project. Decision inference enables the LoRA; native text/image generation disables it and uses the unmodified foundation.

Qev scores supplied answer options through a Jev / TypeSafe-style API (POST /v1/systemone) with real probabilities, not generated JSON text. Compared with the earlier Qwen3.5-0.8B checkpoint (twainsk/qev-0.8b-mlx), this model is 3.4× smaller, answers Snake decisions 2.1× faster and generates chat text about 2× faster on the same Mac, at a measured cost of −2.8 points on the general text subset. The text-only companion is twainsk/qev-230m-mlx.

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                 # downloads and verifies this repository (~1.0 GB) into ~/.qev
qev decide "Card declined twice at checkout" -i "Which team should handle this?" \
  --choice "billing=Invoices and refunds" "technical=Bugs and outages" sales
qev decide --image receipt.jpg -i "Is this a restaurant receipt?" --noul
qev serve                # POST /v1/systemone, /v1/chat/completions, web lab at /

From a source checkout: uv sync --python 3.12 --extra mlx, then uv run hf download twainsk/qev-450m-mlx --local-dir models/qev-450m-mlx and pass --model models/qev-450m-mlx. Requirements: Apple Silicon, macOS 14+, Python 3.12+. Validated with MLX 0.32.2 and mlx-vlm 0.7.1; images use mlx-vlm's numpy LFM2-VL processor, so no PyTorch is needed.

Contents

• qev_config.json — Qev format version 3, family: lfm2, runtime: mlx, modalities ["text", "image"].
• decision_adapters.safetensors — LoRA (rank 64, alpha 128) over q/k/v/out_proj, in_proj and the MLP w1/w2/w3 of all language layers, kept separate from the foundation.
• pointer.safetensors — the 256-dim candidate scorer.
• backbone/ — LiquidAI/LFM2.5-VL-450M at revision fc6221ca…34ba, byte-identical to upstream (bfloat16 safetensors, tokenizer, processor, chat template, license).
• reports/ — evaluation, provenance and MLX-parity summaries; release_manifest.json lists sizes and SHA-256 digests.

Training

Supervised imitation on the Qev decision suite: 17,892 / 1,120 / 1,380 training / calibration / development questions (12,000 Snake plus 5,892 general training rows; English plus four bilingual synthetic rule families). One NVIDIA A100 40 GB, 4 epochs, batch 8 × accumulation 2 (4,476 updates, ~18 min), learning rate 1e-4 cosine, BF16 autocast with FP32 master weights. Foundation parameters stayed frozen; 24.5 M parameters (LoRA + pointer) were trained. Temperature 4.59479341998814 was fitted on the held-out calibration split. A 6-epoch / lr 2e-4 variant scored far worse (59.1% general) and was rejected.

Evaluation

Development set (1,380 questions), PyTorch reference; MLX matches within 2 argmax flips (below).

Closed-loop Snake, 20 seeded games: mean 41.4 food, 0 collisions. Reversed-option probe: 85.1% (−1.0 pt), 7.2% of choices flip. Paired against the Qwen3.5-0.8B checkpoint on identical questions: general −2.8 pt (McNemar p=0.024), Snake parity (p=0.63); against the previous LFM2.5-VL recipe: general +3.2 pt (p=0.013).

Image decisions are zero-shot transfer from text-only decision training: on 500 A-OKVQA validation questions the decision head answered 69.4% with the image (37.2% without; chance 25%), the same as the Qwen3.5 checkpoint's 69.4%. Text calibration is not applied to image questions (temperature 1.0).

MLX vs PyTorch parity on all 1,380 development questions: adapter mode 2 argmax flips, max probability difference 0.051; bf16 2 flips, 0.038; net accuracy unchanged. Apple M4 16 GB: Snake decision p50 98 ms (adapter) / 71 ms (bf16, the Homebrew service default), peak MLX memory ≤1.9 GiB, native chat ≈100 tokens/s in bf16.

Limitations

• The decision adapter was trained on text only; image decisions are measured zero-shot (69.4% A-OKVQA), and image probabilities are uncalibrated. Video input is accepted as ordered frames but was not evaluated for decisions.
• Chinese training coverage is four synthetic rule families; the 50-question Chinese subset is small.
• Option-order sensitivity (7.2% flips) is higher than the Qwen3.5 checkpoint (3.6%).
• 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).