New DeepSeek-V4-Flash-DSpark

Faster performance version of DeepSeek V4 Pro / FLash just dropped.

Do we need to enable speculative config in vllm for this or is it built-in to the model?

The weights incorporate the drafter but new support will be needed to inference engines to actually make use of them

Someone has submitted a pull request on SGlang for DSpark support:

Sweet. I hope we get get to run this on antirez DS4 soon.

DSpark single-stream decode: realistic coding-session harness + τ-focused direction by rafaelcaricio · Pull Request > #1 · rafaelcaricio/spark_vllm_docker

  • DSpark is already paper-grade single-stream (~1.5× MTP-1); the earlier absolute >100 tok/s gate was a phantom caused by the missing MTP-1 baseline.
  • The decode cycle is context-independent (~70 ms, ~55 tok/s flat across 0.5k–65k context). The real coding-session slowdown is entirely τ (draft acceptance) decay under realistic multi-turn coding context (τ 4.3 → ~2.8 as context grows to ~80k). Tverify / MoE-kernel / sparse-MLA work is off the table; MoE tile changes are acceptance-unsafe; KVQ was neutral.