vLLM nightly adds a streaming parser for Qwen3

The vLLM nightly build now includes a new for Qwen3 models. It appears to reduce a problem where stopped in the middle of a reply. It also targets failures in that happened when response chunks were split at awkward boundaries.

Mid-reply stops are especially disruptive when the model is used in that need several steps to finish a task. In limited testing, the same stopping problem did not appear again.

Key points

  • vLLM nightly has a new for Qwen3 models.
  • The change appears to reduce mid-reply stopping with .
  • It also aims to fix streaming tool call failures caused by response chunk boundaries.
  • More stable can mean fewer retries and lower token waste.
  • The result still needs practical testing before production use.
Read original