Qwen3.6-27B-MXFP8-MTP

Qwen3.6-27B (dense) quantized to native MXFP8 for Apple Silicon, with the vision tower and the native Multi-Token-Prediction head preserved and enabled.

Source	Qwen/Qwen3.6-27B
License	Apache-2.0, inherited from upstream
Format	MXFP8 (mx.quantize, affine, group_size=32)
Architecture	qwen3_5 dense — 64 layers, hybrid GatedDeltaNet + full attention, hidden 5120
Modality	image + video + text
Context	262,144
Bundle size	27.1 GB
MTP	native head preserved, enabled (num_nextn_predict_layers=1)

Quantization

8-bit affine linears via MLX-native mx.quantize (mode="mxfp8", group_size=32). Norms, hybrid-attention control tensors and the full vision tower are kept in fp16 passthrough (693 passthrough tensors). MTP linears are quantized to MXFP8; MTP norm/control tensors stay fp16.

Multi-Token Prediction

This bundle keeps Qwen3.6's native MTP module and runs it as a self-speculative draft head: the MTP head proposes tokens that the main model verifies in a single pass, so decoded output stays bit-identical to plain autoregressive decoding — only faster.

Recorded on an M5 Max (vMLX runtime, 96-token deterministic prompt, output verified equal to baseline at every depth):

Draft depth	tok/s	Speedup
Baseline (MTP off)	15.8	1.00×
D1	24.7	1.56×
D2	28.8	1.82×
D3 (default)	28.9	1.83×

With vMLX prefix/KV cache layers enabled the speedup holds — a recorded cache-on A/B measured 15.5 → 28.2 tok/s (1.81×).

Absolute tok/s depends on free memory and system load. The speedup ratio — baseline vs. MTP measured back-to-back under identical conditions — is the stable figure.

Vision, MTP and caching together

These bundles run image/video input, native MTP speculative decode and prefix/KV caching in the same session — a combination not every MTP-enabled Qwen build exposes. A recorded VL probe (2026-05-16) confirms a color identification image prompt returns the correct answer through the combined MTP + VL runtime.

Loading

Loads via stock MLX tooling on Apple Silicon — the mxfp8 weights are native mx.quantize affine, no JANG runtime required for the core model.

from mlx_vlm import load, generate
model, processor = load("OsaurusAI/Qwen3.6-27B-MXFP8-MTP")

The MTP draft path is exercised by an MTP-aware runtime (vMLX); other runtimes load and decode the main model normally and ignore the MTP head.

Variants

Variant	Arch	Format	Size	Best MTP speedup
Qwen3.6-27B-MXFP4-MTP	dense	mxfp4	14.4 GB	1.85× (D2)
Qwen3.6-27B-MXFP8-MTP (this)	dense	mxfp8	27.1 GB	1.83× (D3)
Qwen3.6-35B-A3B-MXFP4-MTP	MoE	mxfp4	21.5 GB	1.56× (D3)
Qwen3.6-35B-A3B-MXFP8-MTP	MoE	mxfp8	35.0 GB	1.71× (D3)

Credits

• Quantization toolchain: JANG by Jinho Jang <eric@osaurus.ai>
• Base model: Qwen3.6-27B by Qwen