Modelpublic

barozp/Qwen3.8-Whittle-MoE-27B-MLX-4bit

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

5likes2.4kdownloads

Model Card

Qwen3.8-Whittle-MoE-27B-MLX-4bit

MLX 4-bit conversion of [logic65/Qwen3.8-Whittle-MoE-27B-A17.8B](https://huggingface.co/logic65/Qwen3.8-Whittle-MoE-27B-A17.8B) -- a post-hoc MoE carved from Qwen/Qwen3.8-27B (64 routed expert slivers + 1 shared expert per layer, top-16 routing, 17.8B active of 27B total, router-healed + anti-loop trained, v2.1) -- for Apple Silicon. Sibling formats:

—bf16 safetensors source (v2.1): logic65/Qwen3.8-Whittle-MoE-27B-A17.8B
—GGUF quantizations: logic65/Qwen3.8-Whittle-MoE-27B-A17.8B-GGUF

Converted straight from the v2.1 bf16 safetensors weights (not from any GGUF quant). Note: the source ships without an MTP head (mtp_layers: 0) -- this MLX build is autoregressive only; MTPLX speculative decoding does not apply.

Conversion details

—Bits: 4 (group_size=64, affine)
—Command: python -m mlx_lm.convert --hf-path logic65/Qwen3.8-Whittle-MoE-27B-A17.8B --mlx-path <out> --quantize --q-bits 4 --q-group-size 64
—Tooling: mlx 0.32.1 · mlx-lm 0.31.3 · hub 1.27.0; source commit 08741134dbc3
—Converter: mlx-lm
—Environment: Google Colab G4 (RTX PRO 6000 Blackwell 96 GB), MLX cuda12 backend
—Full conversion log ships in this repo as conversion_log.txt

Usage

bash

pip install -U mlx-lm

# text-only chat
python -m mlx_lm.generate \
  --model barozp/Qwen3.8-Whittle-MoE-27B-MLX-4bit \
  --prompt "Explain the Monty Hall problem briefly." \
  --max-tokens 512

# OpenAI-compatible local server
python -m mlx_lm.server \
  --model barozp/Qwen3.8-Whittle-MoE-27B-MLX-4bit

About the source model

A post-hoc MoE carved from Qwen3.8-27B: each layer's dense FFN (17408) was partitioned into 64 routed expert slivers of width 192 + one shared expert of width 5120; a small router picks 16 of 64 slivers per token (17.8B active / 27B total). Freshly carved it scored 4/39 on the knowledge battery; training only the 64 routers recovered 27/39. The v2.1 lineage adds anti-loop training (245 complete teacher answers with top-32 logprobs) and a balancing epoch: single-turn loop rate 8% (from 69%), structured-output failures 22% (from ~75%), knowledge battery 28/39. Full method and failure log: WHITTLE_FINDINGS.md.

Source chain

Qwen/Qwen3.8-27B (dense base) -> logic65/Qwen3.8-Whittle-MoE-27B-A17.8B (post-hoc MoE carve + router healing + v2.1 anti-loop) -> this repo (MLX 4-bit conversion)