Bring your own drafter

# a measured pair: the drafter resolves by itself
mlx-dspark generate --model mlx-community/Qwen3-14B-8bit --prompt "Explain how rainbows form."

# anything else: name the drafter; the machinery is identical from here on
mlx-dspark generate --model mlx-community/Qwen3-32B-8bit \
  --drafter deepseek-ai/dspark_qwen3_32b_block7 --prompt "Explain how rainbows form."

What runs, and what doesn't#

New DSpark and DFlash drafters keep landing on Hugging Face in several different packagings. The loaders recognize these, and refuse anything incompatible with an error naming the reason, never a silent mis-load.

One measured example outside the registry: makora-ai/gemma4-26b-a4b-dspark on mlx-community/gemma-4-26b-a4b-it-8bit (Google's 26B mixture of experts) runs at 1.27× with --max-draft 2 --no-lookup-drafts: 1.38× on code, 1.37× on math, 1.06× on chat. It isn't registered while the ratio is under review.

Which models can be drafted for#

• Any dense mlx-lm text model routes automatically. On load, a one-time probe checks that the hidden-state tap the drafter reads reproduces the model's own forward pass, and fails loudly if the model's family needs dedicated support.
• Hybrid and recurrent families already supported: Qwen3.5/3.6/3.8 (linear attention), Nemotron-H (Mamba-2), LFM2 (short convolutions), Nanbeige (looped depth), and Gemma-4 and Muse through mlx-vlm. Their recurrent state can't be trimmed like a KV cache, so each verify round's inputs are recorded and the state is rebuilt exactly at the accept point.
• Any model at all runs drafter-free with --mode lookup, or --mode auto.

Getting a good pairing#

• Use the matched instruct model the drafter was trained against. A base model drops acceptance sharply.
• Match the model's precision to the drafter's training. A drafter trained against the bf16 model does best on the 8-bit build, one trained against a 4-bit model on the 4-bit build. (Better training data can beat this rule: Red Hat's bf16-trained Qwen3.8 head wins on the 4-bit model too.)
• Drafter precision doesn't matter for acceptance. Drafters load 4-bit by default (--drafter-bits changes it), because that's the fastest to run.
• Very small models aren't worth a drafter. A Qwen3.5-0.8B DSpark head runs correctly and losslessly at 0.96×: the model already decodes at 215 tok/s, and a draft round costs nearly what it saves. Run those models plainly.

Measure it#

mlx-dspark benchmark --model <model> --drafter <drafter> --caps 2,4,7,auto --trials 3 --json result.json

The registered pairs accept between 2.5 and 5.5 tokens per round. Acceptance stuck near 1.2 means something is mismatched: a base model instead of the instruct one, the wrong quantization, or a packaging detail the loader doesn't know yet. If you've measured a pair that isn't listed, the device-stamped JSON is exactly what's needed to add it. Please open an issue with it.

Remote code#

Drafters load their weights one-to-one and never import code, so a drafter whose config carries an auto_map pointer needs nothing special. Models that ship their own Python (model_file or auto_map in config.json) are refused by default, because the loaders would otherwise import and run it. Opt in with --trust-remote-code or MLX_DSPARK_TRUST_REMOTE_CODE=1 once you trust the repo.