JetFlow: Breaking the Scaling Ceiling of Speculative Decoding with Parallel Tree Drafting

In three linesJetFlow improves speculative decoding by combining parallel drafting efficiency with branch-wise causal conditioning. On H100 GPUs, it achieves 9.64x speedup on MATH-500 and 4.58x on open-ended conversations, outperforming existing tree-based methods on dense and MoE Qwen3 models.

Summary generated by Claude — human-verified