Algorithm Decision Guide
Speculators currently supports six speculative decoding algorithms: Eagle-3, P-EAGLE, DFlash, DFlash2, DSpark, and MTP. All are lossless -- they produce output from the same distribution as the target model.
How They Differ
Eagle-3 predicts draft tokens autoregressively, one at a time.
P-EAGLE extends Eagle-3 with parallel multi-token prediction across multiple depths, using COD sampling for memory-efficient training.
DFlash predicts all draft tokens in a single forward pass using block-based prediction with anchor points.
DFlash2 extends DFlash with local dynamic convolutions and a low-rank candidate selector. Its training objective is experimental and follows the upstream DFlash2 candidate-set contract.
DSpark extends DFlash with a Markov head, so positions inside a block depend on earlier ones, plus a confidence head that predicts per-position acceptance.
MTP finetunes the model's native multi-token prediction head on domain-specific data. Unlike the other algorithms, MTP does not train from scratch -- it starts from pre-existing MTP layers and is only available for models with native MTP support.
Eagle-3, P-EAGLE, DFlash, and DSpark can be paired with any supported verifier model (including quantized variants). DFlash2 requires the verifier's full vocabulary because its selector operates on the complete vocabulary. MTP requires a model with native MTP layers (e.g. Qwen3-Next, Qwen3.5).
Current Support
| Eagle-3 | P-EAGLE | DFlash | DFlash2 | DSpark | MTP | |
|---|---|---|---|---|---|---|
| Draft layers | Llama-style | Llama-style | Qwen3-style | Qwen3-style | Qwen3-style | Native MTP layers |
| Verifier models | Any supported | Any supported | Any supported | Full vocabulary | Any supported | Models with native MTP only |
| Training mode | From scratch | From scratch | From scratch | Experimental | From scratch | Finetune existing MTP head |
| Speculators | Mature | Newer, growing fast | Newer, growing fast | Experimental | Newer, growing fast | Newer, growing fast |
| vLLM | Mature | Newer, growing fast | Newer, growing fast | Native/experimental | Newer, growing fast | Newer, growing fast |
Eagle-3 has been available longer and has broader support in both Speculators and vLLM. The others were added more recently and support is improving rapidly; DFlash2 remains experimental.
Which Should I Use?
If you're unsure, start with Eagle-3 -- it has the most mature tooling and documentation. If you want parallel multi-token prediction with an Eagle-3-based architecture, try P-EAGLE. If you want to experiment with DFlash's single-forward-pass block prediction approach, the training workflow is the same. DFlash2 adds local convolution and candidate reranking, while DSpark adds intra-block dependencies and confidence scheduling. If your model already has native MTP layers (e.g. Qwen3-Next, Qwen3.5), MTP finetuning lets you improve the existing MTP head on domain-specific data without training a separate draft model.
For more details on each algorithm, see: