DFlash
A speculative decoding acceleration paradigm in which an independent small draft model proposes blocks of candidate tokens in batches, and the main model then verifies them in parallel in a single forward pass, accepting matching portions and resampling non-matching portions. It can achieve a relatively high speedup ratio.
Timeline (last 90 days)
In the MLX backend changes, when the draft model configuration requires an independent output head, the draft's own output head is loaded while retaining the target head path used by existing DFlash models.
The MLX backend of Ollama v0.40.10 supports speculative decoding using a DFlash draft model for the Gemma 4 target model.
On the MLX inference path, when the draft model's configuration requires it, v0.40.10-rc0 loads the draft's own output head while retaining the target-head path used by existing DFlash models.
When speculative decoding is configured properly, token generation speed can improve by about 1.5x (according to vLLM); Omer personally tested Qwen3 35B MoE with a DFlash drafter in his home lab and measured a 2x to 3x improvement, close to 3x in coding scenarios and about 1.5x in browser automation.
The vLLM team worked with Reflection AI and Inferact to jointly optimize the serving path, including fused kernels and DFlash.
On a single H100 and at similar measured accuracy, QwenDLM-4B + ORL reaches 1,298 tokens/s, DFlash 746 tokens/s, and EAGLE3 405 tokens/s (end-to-end system comparison: ORL runs on dlmrt, speculative decoding runs on vLLM).
Nemotron 3.5 Bolt comes with a LiLiCorr drafter for speculative decoding, implemented via DFlash, with 7 speculative tokens by default in the vLLM configuration.
Qwen3.8-27B是稠密模型，全部270亿参数参与每个token的计算，依靠MTP头预测后续token并在单次前向中验证；Ornith 1.5采用MoE架构，每token仅激活约30亿参数，配合DFlash由小型草稿模型提议token块再由主模型校验。
MiMo-V2.6 has a built-in DFlash speculative decoding mechanism that can draft 7 tokens per step.
DFlash and DSpark have made targeted optimizations in draft generation efficiency and hardware adaptation, and may outperform classic solutions under specific models and workloads.
1 more timeline events
All Facts (12)
MTP与DFlash在架构上有本质不同，但都属于投机解码技术谱系
85%UnverifiedIn the MLX backend changes, when the draft model configuration requires an independent output head, the draft's own output head is loaded while retaining the target head path used by existing DFlash models.
50%UnverifiedOn the MLX inference path, when the draft model's configuration requires it, v0.40.10-rc0 loads the draft's own output head while retaining the target-head path used by existing DFlash models.
50%UnverifiedWhen speculative decoding is configured properly, token generation speed can improve by about 1.5x (according to vLLM); Omer personally tested Qwen3 35B MoE with a DFlash drafter in his home lab and measured a 2x to 3x improvement, close to 3x in coding scenarios and about 1.5x in browser automation.
50%UnverifiedThe vLLM team worked with Reflection AI and Inferact to jointly optimize the serving path, including fused kernels and DFlash.
50%UnverifiedOn a single H100 and at similar measured accuracy, QwenDLM-4B + ORL reaches 1,298 tokens/s, DFlash 746 tokens/s, and EAGLE3 405 tokens/s (end-to-end system comparison: ORL runs on dlmrt, speculative decoding runs on vLLM).
50%UnverifiedNemotron 3.5 Bolt comes with a LiLiCorr drafter for speculative decoding, implemented via DFlash, with 7 speculative tokens by default in the vLLM configuration.
50%UnverifiedQwen3.8-27B是稠密模型，全部270亿参数参与每个token的计算，依靠MTP头预测后续token并在单次前向中验证；Ornith 1.5采用MoE架构，每token仅激活约30亿参数，配合DFlash由小型草稿模型提议token块再由主模型校验。
50%UnverifiedMiMo-V2.6 has a built-in DFlash speculative decoding mechanism that can draft 7 tokens per step.
50%UnverifiedDFlash and DSpark have made targeted optimizations in draft generation efficiency and hardware adaptation, and may outperform classic solutions under specific models and workloads.
50%UnverifiedSpeculative decoding has spawned multiple variants including Medusa (multi-head prediction), EAGLE (generating drafts using hidden states from early layers of the main model), and DFlash (hybrid layout optimization).
50%UnverifiedThe MLX backend of Ollama v0.40.10 supports speculative decoding using a DFlash draft model for the Gemma 4 target model.
50%