A new speculative decoding method called DFlash, now supported in vLLM v0.25.0 for CPUs, boosts average token generation throughput to 3.92 times the autoregressive baseline, cutting per-token costs by roughly 74%. DFlash pairs a lightweight draft model with a larger target model, predicting blocks of tokens in one pass and verifying them together. This reduces memory-bound, sequential decoding steps and better utilizes Intel Xeon 6 CPU compute engines. Benchmarks on Qwen3.5-9B across coding, math, and conversational datasets show strong performance gains, especially for structured prompts. Developers can reproduce results via published configurations and explore compatible drafter models on Hugging Face.
This update represents a notable development in the Ai sector. Organizations and founders tracking this space should evaluate potential strategic and technical implications on their operations.