license: apache-2.0 base_model: AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16 library_name: mlx tags:

mlx
mlx-vlm
qwen3_5_mtp
qwen3_5
qwen3.8
mtp
multi-token-prediction
speculative-decoding
self-speculation
draft-model
apple-silicon

Qwen3.8-27B-AEON MTP Drafter (MLX)

The native multi-token-prediction head of AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16 (revision 8f76e82), split into the standalone MLX drafter format that mlx-vlm expects for mtp speculative decoding. 15 tensors, BF16, 810 MB, block_size: 3.

AEON's BF16 grafts the MTP head back from stock Qwen3.8 (15 tensors, hash-matched), so this drafter is the original Qwen-trained head — not retrained.

Companion to VisualInference/Qwen3.8-27B-AEON-Ultimate-Uncensored-Multimodal-MLX-6bit; works with any MLX quantization of the same base.

Status (2026-09-10)

This drafter still works as documented, and --draft-block-size 3 is still the right setting for it. Two measured facts since publication, for anyone choosing a speculative-decoding setup on a Mac:

On a 6-bit Qwen3.8-27B body at 24-27k tokens of agent context, mlx-vlm with a native MTP drafter measured 6-7 tok/s end to end, while mlx-dspark with incoai/Qwen3.8-27B-DFlash2 measured 15-19 tok/s on the same body and prompts. Both are lossless. At short context the gap is smaller (this card's numbers were measured at up to 13k).
DFlash2 is trained on stock Qwen3.8 and works unchanged with any abliterated Qwen3.8-27B body, including AEON's; it changes speed, never output.

The lane recipe, launcher settings and patches: qwen38-27b-mlx-agent-lane.

How it was made

python -m mlx_vlm.speculative.drafters.qwen3_5_mtp.split \
  --model AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16 \
  --output Qwen3.8-27B-AEON-Ultimate-Uncensored-MLX-MTP-Drafter

(Use the official split tool — hand-extracting mtp.* tensors produces a broken drafter with 0% acceptance: the tool applies the RMSNorm weight-convention shift and stamps format: mlx metadata.)

Usage

python -m mlx_vlm generate \
  --model VisualInference/Qwen3.8-27B-AEON-Ultimate-Uncensored-Multimodal-MLX-6bit \
  --draft-model VisualInference/Qwen3.8-27B-AEON-Ultimate-Uncensored-MLX-MTP-Drafter \
  --draft-kind mtp --draft-block-size 3 \
  --prompt "..."

Speculative decoding is lossless — rejected drafts fall back to the target model's own tokens.

Measured effect

Test machine: Mac mini M4 Pro, 48 GB, with the companion 6-bit target model. Speedups carry across chips since the drafter's relative overhead is small.

Workload	Serial	With drafter
Coding (temp 0.2, block 4)	11.4 tok/s	21.5 tok/s (1.90×)
Doc QA @ 13k ctx (block 3)	~11 tok/s	16.9 tok/s
Creative prose (temp 0.7, block 3)	11.4 tok/s	15.9 tok/s (1.4×)

Block size 3 is the all-round sweet spot; block ≥5 regresses (draft accuracy decays past 2–3 tokens). Acceptance ~46% on open prose, higher on structured output.

Qwen3.8-27B-AEON MTP Drafter (MLX)

AEON's BF16 grafts the MTP head back from stock Qwen3.8 (15 tensors, hash-matched), so this drafter is the original Qwen-trained head — not retrained.

Status (2026-09-10)

This drafter still works as documented, and --draft-block-size 3 is still the right setting for it. Two measured facts since publication, for anyone choosing a speculative-decoding setup on a Mac:

On a 6-bit Qwen3.8-27B body at 24-27k tokens of agent context, mlx-vlm with a native MTP drafter measured 6-7 tok/s end to end, while mlx-dspark with incoai/Qwen3.8-27B-DFlash2 measured 15-19 tok/s on the same body and prompts. Both are lossless. At short context the gap is smaller (this card's numbers were measured at up to 13k).

DFlash2 is trained on stock Qwen3.8 and works unchanged with any abliterated Qwen3.8-27B body, including AEON's; it changes speed, never output.

The lane recipe, launcher settings and patches: qwen38-27b-mlx-agent-lane.

How it was made

python -m mlx_vlm.speculative.drafters.qwen3_5_mtp.split \ --model AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16 \ --output Qwen3.8-27B-AEON-Ultimate-Uncensored-MLX-MTP-Drafter

(Use the official split tool — hand-extracting mtp.* tensors produces a broken drafter with 0% acceptance: the tool applies the RMSNorm weight-convention shift and stamps format: mlx metadata.)

Usage

python -m mlx_vlm generate \ --model VisualInference/Qwen3.8-27B-AEON-Ultimate-Uncensored-Multimodal-MLX-6bit \ --draft-model VisualInference/Qwen3.8-27B-AEON-Ultimate-Uncensored-MLX-MTP-Drafter \ --draft-kind mtp --draft-block-size 3 \ --prompt "..."

Speculative decoding is lossless — rejected drafts fall back to the target model's own tokens.

Measured effect

Test machine: Mac mini M4 Pro, 48 GB, with the companion 6-bit target model. Speedups carry across chips since the drafter's relative overhead is small.

Workload

Serial

With drafter

Coding (temp 0.2, block 4)

11.4 tok/s

21.5 tok/s (1.90×)

Doc QA @ 13k ctx (block 3)

~11 tok/s

16.9 tok/s

Creative prose (temp 0.7, block 3)

11.4 tok/s

15.9 tok/s (1.4×)

Block size 3 is the all-round sweet spot; block ≥5 regresses (draft accuracy decays past 2–3 tokens). Acceptance ~46% on open prose, higher on structured output.