Qwen3.5-9B MTPLX Optimized Speed

Fast local 9B inference for Apple Silicon, packaged for MTPLX native Multi-Token-Prediction speculative decoding.

This is the 9B speed checkpoint: a 6-bit MLX body with BF16 MTP heads, tuned as the stronger small-Mac option above the 4B release.

Run It

brew install youssofal/mtplx/mtplx
mtplx start
mtplx run "hello" --model Youssofal/Qwen3.5-9B-MTPLX-Optimized-Speed

For an OpenAI-compatible local server:

mtplx serve --model Youssofal/Qwen3.5-9B-MTPLX-Optimized-Speed --profile sustained --max --port 8000 --no-stats-footer

Why This Exists

MTPLX uses the model's own MTP heads to generate draft tokens, then verifies those tokens with the main model. When the draft heads are well-matched, you get higher throughput without running a separate drafter model.

MTPLX reads mtplx_runtime.json and selects the measured defaults automatically.

Recommended Runtime Defaults

Setting	Value
Backend	`qwen3-next-mtp`
Default depth	`D2`
Target sampler	`temp=0.60`, `top_p=0.95`, `top_k=20`
Draft sampler	`temp=0.60`, `top_p=0.95`, `top_k=20`
Profile	`sustained`
Benchmark fan mode	`max`

Performance

Measured in MTPLX on Apple Silicon using the release runtime path.

Mode	TPS	Verify time	Acceptance
AR baseline	64.96	-	-
D1 comparison	92.87	6.83s	0.9120
D2 promoted default	101.32	4.13s	0.9398, 0.8102
D3 comparison	96.30	4.58s	0.9278, 0.7732, 0.6443

Model Build

Component	Format
Main body	6-bit MLX affine, group size 64
MTP heads	BF16 native MTP sidecar
Sampler	target and draft temp=0.60, top_p=0.95, top_k=20

This is not a full-precision checkpoint. It is built for fast local use on Apple Silicon through MTPLX.

Files

model-*.safetensors: MLX body shards
mtp.safetensors: MTP sidecar
mtplx_runtime.json: MTPLX runtime contract and measured defaults
MTPLX_PUBLISH_MANIFEST.json: file sizes and artifact metadata
tokenizer and config files for local loading