Qwen3.8-27B-MTPLX-bf16

MTP weights kept. This is a full-precision MLX build — no quantization at all. The multi-token-prediction head is preserved at bf16 (mtp_policy: keep_bf16), same as the body.

A full-precision, unquantized MLX build of Qwen/Qwen3.8-27B for Apple Silicon, produced with mtplx forge. This is not a repackaging of the original PyTorch/safetensors bf16 release — it's converted into native MLX format for direct loading on Apple Silicon, with the MTP head intact so speculative decoding works out of the box.

Why this build exists

The original Qwen/Qwen3.8-27B weights are safetensors in HF/PyTorch format. Loading those directly through generic MLX conversion paths risks losing the model's 15 MTP tensors (see the 4-bit/8-bit sibling cards for the full explanation of why). This build applies the same mtp_policy: keep_bf16 forge recipe as the quantized variants, but with body_bits: 0 — meaning the trunk is not quantized at all. It exists for anyone who wants zero quantization-related quality loss and has the memory to spare.

Recipe

Parameter	Value
Body precision	bf16 (unquantized)
MTP head	kept at bf16
Source	Qwen/Qwen3.8-27B (bf16 native)

Requirements

• Apple Silicon Mac (M-series) with 64GB+ unified memory recommended — full bf16 weights are ~55GB on disk/in memory before KV cache.
• mtplx — required to use the native MTP speculative-decode path.

Usage

mtplx quickstart --model johninthepool/Qwen3.8-27B-MTPLX-bf16 --port 8020 \
  --reasoning off --paged-kv-quantization q8

Which variant should you use?

• This one (bf16): maximum fidelity, largest footprint. Use if you have the RAM and want zero quantization loss.
• 8-bit: near-lossless, roughly half the footprint. Best default for most 64GB+ machines.
• 4-bit: smallest footprint, best throughput, small quality tradeoff.

Provenance

Converted directly from Qwen/Qwen3.8-27B to MLX format with no fine-tuning, distillation, or precision changes beyond format conversion. Output is expected to be numerically equivalent to the source model.