APUS-Reflex-9B-MLX-4bit

APUS Reflex 9B-MLX-4bit

Branding note: This model family is now APUS Reflex. Some weight filenames, script/module names, served-model aliases, and the separate evaluation dataset ID retain their historical identifiers for download and runtime compatibility; these do not indicate different model versions.

English | 中文 · Source model · Collection · GGUF collection · MLX collection · MLX-8bit · GGUF / Ollama

MLX weights (mixed 4/6-bit affine (mlx-lm mixed_4_6), group size 64) of APUS-Reflex-9B for Apple Silicon Macs (mlx-lm, LM Studio).

APUS Reflex is a decision model: each request supplies a state, an instruction and 2–16 candidates, and the model scores candidate labels A–P. It is not a chat model.

Quick start

pip install mlx-lm
hf download apus-ailab/APUS-Reflex-9B-MLX-4bit --local-dir ./apus-reflex
python ./apus-reflex/examples/openjev_mlx.py --model ./apus-reflex

examples/openjev_mlx.py renders prompts with openjev_contracts.py (the training contract) and returns the exact candidate distribution.

Parity

Frozen80 with identical prompt tokens, compared with the HF BF16 release (full depth, 68/80 · 85.00%):

Converted and scored with MLX on Linux (CUDA); the files are platform-independent and load unchanged on Apple Silicon (for 4B, the same kind of file gave identical decisions on CUDA and Metal). Peak memory on Frozen80 was 7.42 GB; plan for a Mac with at least 16 GB of unified memory. Frozen80 is a reused development panel, not a blind benchmark. Per-question rows (candidate probabilities, choice, correctness; join with Frozen80 by panel_index): evaluation/per-question/.

Conversion

• mlx-lm 0.31.3 / mlx 0.32.2; mixed 4/6-bit affine (mlx-lm mixed_4_6), group size 64.
• 6-bit lm_head and v_proj/down_proj in sensitive layers, 4-bit elsewhere (the MLX analogue of Q4_K_M); MoE router gates stay 8-bit (recipe). Calibrated DWQ/GPTQ were tried and were not practical for this hybrid-attention model on the conversion hardware.
• GDN A_log and linear_attn.norm.weight keep their source precision (FP32 in the 35B release).
• Full depth only, text only, probabilities not calibrated.

License

Apache-2.0, inherited from the source model; see LICENSE. Base model: Qwen/Qwen3.5-9B.

Authors: gumpcheng (xDAN2099), zhangxu, APUS AI-LAB.