FancieF/Qwythos-27B-v1-MLX-MTP

Qwythos-27B-v1 · MLX MTP Drafter

Standalone Multi-Token Prediction (MTP) head split from empero-ai/Qwythos-27B-v1.

Not a full LLM. Load only as --draft-model next to an MLX VLM target.

Which repo should I download?

Repo	Role
VLM 4-bit	Main model (recommended)
VLM BF16	Main model (high RAM)
This MTP drafter	Optional decode speed-up

mlx_vlm.convert drops mtp.*. This repo puts them back as model_type: qwen3_5_mtp (block_size=4, ~829MB).

Install

pip install -U mlx-vlm

Usage (copy-paste)

mlx_vlm.generate \
  --model FancieF/Qwythos-27B-v1-MLX-VLM-4bit \
  --draft-model FancieF/Qwythos-27B-v1-MLX-MTP \
  --draft-kind mtp \
  --draft-block-size 4 \
  --prompt "Explain gated linear attention in three sentences." \
  --thinking-mode disabled \
  --temperature 0.6

With image (prefill on main model; MTP helps the text tail):

mlx_vlm.generate \
  --model FancieF/Qwythos-27B-v1-MLX-VLM-4bit \
  --draft-model FancieF/Qwythos-27B-v1-MLX-MTP \
  --draft-kind mtp \
  --draft-block-size 4 \
  --image ./photo.jpg \
  --prompt "Describe this image." \
  --thinking-mode disabled

Server:

mlx_vlm.server \
  --model FancieF/Qwythos-27B-v1-MLX-VLM-4bit \
  --draft-model FancieF/Qwythos-27B-v1-MLX-MTP \
  --draft-kind mtp \
  --draft-block-size 4

LM Studio note

Do not expect this file to work as LM Studio’s generic Draft Model. For LM Studio MTP, use official GGUF with MTP baked in: empero-ai/Qwythos-27B-v1-GGUF (*MTP*Q4_K_M*) + enable MTP Speculative Decoding.

How it was produced

python -m mlx_vlm.speculative.drafters.qwen3_5_mtp.split \
  --model /path/to/Qwythos-27B-v1 \
  --output ./Qwythos-27B-MTP \
  --block-size 4

License & attribution

Derived from empero-ai/Qwythos-27B-v1 (Apache-2.0). Unofficial MLX drafter for Apple Silicon.

Links

Main VLM 4-bit	FancieF/Qwythos-27B-v1-MLX-VLM-4bit
VLM BF16	FancieF/Qwythos-27B-v1-MLX-VLM-bf16
Base	empero-ai/Qwythos-27B-v1
ModelScope (ZH)	Fancie/Qwythos-27B-v1-MLX-MTP