Which of the following are currently supported with PFlash + DFlash/DDTree?

[TomLucidor]

• New release from Gemma 4 https://huggingface.co/z-lab/gemma4-12B-it-DFlash
• https://huggingface.co/z-lab/gemma-4-31B-it-DFlash
• https://huggingface.co/z-lab/gemma-4-26B-A4B-it-DFlash
• Small models https://huggingface.co/z-lab/Qwen3.5-4B-DFlash

Already supported should be https://huggingface.co/z-lab/Qwen3.5-27B-DFlash and https://huggingface.co/z-lab/Qwen3.5-35B-A3B-DFlash and by extension https://huggingface.co/z-lab/Qwen3.6-27B-DFlash and https://huggingface.co/z-lab/Qwen3.6-35B-A3B-DFlash according to the documentation
https://huggingface.co/z-lab/Qwen3.5-9B-DFlash is also covered when doing DDTree testing. But small Gemma models are not yet supported for most things?

[raullenchai]

Thanks for asking. I checked the current implementation and alias registry. The short answer is: support is not inherited by model family, and PFlash is not stacked with the current DFlash/DDTree servers.

Current validated matrix:

Target	PFlash	DFlash	DDTree
Qwen3.5 4B	yes (verified aliases)	no	no
Qwen3.5 9B 8-bit	yes separately	no	yes, with z-lab/Qwen3.5-9B-DFlash
Qwen3.5 27B 8-bit	yes separately	yes, with z-lab/Qwen3.5-27B-DFlash	no
Qwen3.5/3.6 35B-A3B	yes for the verified aliases	no	no
Qwen3.6 27B	yes for the verified aliases	no	no
Gemma 4 12B / 31B / 26B-A4B	not validated/auto-enabled	no	no

A few important details:

• The only DFlash-enabled target today is qwen3.5-27b-8bit.
• The only DDTree-enabled target today is qwen3.5-9b-8bit.
• The 4-bit target aliases are deliberately rejected by DFlash/DDTree. These lanes currently require a validated target at 8-bit or higher.
• Qwen3.5/3.6 35B-A3B and Gemma 4 26B-A4B are MoE targets. The current DFlash gate rejects MoE, and the DDTree MVP is validated only for a dense Qwen target.
• A matching-looking drafter repository does not make a target supported automatically. Each exact target/drafter pair must be benchmarked and explicitly enabled in aliases.json. Therefore the Qwen3.6 models are not supported “by extension” from Qwen3.5.
• Gemma 4 models themselves are supported for normal serving, but the new Gemma DFlash checkpoints are not wired into Rapid-MLX's DFlash or DDTree runtime yet. PFlash is also not validated for these aliases; in the default multimodal lane it is unavailable.

Also, “PFlash + DFlash/DDTree” should not be read as a combined mode. DFlash and DDTree use dedicated single-user server paths that bypass BatchedEngine, where PFlash runs. DFlash explicitly ignores --pflash; DDTree likewise does not use the PFlash path.

You can inspect the live decision for an installed version with:

rapid-mlx info qwen3.5-27b-8bit
rapid-mlx info qwen3.5-9b-8bit
rapid-mlx models

So, for the checkpoints listed in the question:

• gemma4-12B-it-DFlash: not currently supported by DFlash/DDTree
• gemma-4-31B-it-DFlash: not currently supported by DFlash/DDTree
• gemma-4-26B-A4B-it-DFlash: not currently supported by DFlash/DDTree
• Qwen3.5-4B-DFlash: not currently supported by DFlash/DDTree
• Qwen3.5-9B-DFlash: supported as the drafter for DDTree with the 9B 8-bit target
• Qwen3.5-27B-DFlash: supported as the drafter for DFlash with the 27B 8-bit target
• Qwen3.6 DFlash checkpoints: not currently enabled

[raullenchai]

An update to the accepted answer for v0.16.3:

• Verified DFlash aliases: qwen3.5-27b-8bit, qwen3.8-27b-tensorfold, bonsai2-27b-tensorfold and muse-glimmer-30b-8bit.
• DDTree: still only qwen3.5-9b-8bit, with z-lab/Qwen3.5-9B-DFlash.
• 4-bit and unverified dense targets are no longer hard-rejected. You can opt in with --speculative-config '{"method":"dflash","model":"<drafter>"}'. After structural checks pass, the pair runs with an experimental warning (see docs/reference/configuration.md).
• Unchanged: MoE targets are still rejected by DFlash. The Gemma 4 DFlash drafters are not a verified pairing. PFlash is still not combined with DFlash or DDTree, and DFlash ignores --pflash.

rapid-mlx info <alias> shows the live decision for your installed version. Closing, since the original question is answered.