DFlash 2 is a speculative decoding technique that improves draft-token generation speed during LLM inference. The core idea: instead of guessing one token at a time (autoregressive), DFlash predicts an entire block in parallel. DFlash 2 adds two components—a lightweight path selector that picks coherent token sequences from candidate lists using pairwise scoring (only 2M parameters), and short convolutions that preserve local dependencies across the block. The result: 20–25% more accepted tokens per verification pass with ~1% latency overhead. Concretely, on Qwen3.8-27B, this delivers 2.7–3.4× throughput versus autoregressive decoding at batch size 1, and on Meta's Muse Glimmer, 3.1–4.6×. The drafters are available now on Hugging Face and integrated into SGLang, vLLM, and llama.cpp. The practical win: agents that read and call tools for hours burn tokens at scale; anything cutting inference time by ~3× directly cuts serving cost and latency.