Inference has become the bottleneck of the agent era, with agents consuming tokens at unprecedented rates. Inco AI is building the inference stack for this new token economy. The original DFlash, released in January, now runs across SGLang, vLLM, TensorRT-LLM, and llama.cpp, achieving up to 15x throughput on NVIDIA Blackwell GPUs and 3x more tokens per second on TPUs.

DFlash introduced one-pass speculative drafting, predicting entire token blocks in parallel rather than autoregressively. DFlash 2 builds on this by recovering over 20% more output per verification pass with only 1% added latency. The approach uses a lightweight path selector that scores adjacent token pairs instead of relying on costly sequential corrections.

Qwen3-8-27B DFlash 2 drafters are available today, enabling SGLang to serve at 2.7-3.4x the throughput of autoregressive decoding. The selector keeps top 16 candidates per position and traces coherent paths through them using low-rank bilinear attention, staying fully parallel with minimal overhead.

DFlash 2 already runs in mainstream inference engines including SGLang, vLLM, llama.cpp, and Ollama, with installation instructions provided for each platform.