Skip to content

v0.6.7

Choose a tag to compare

View all tags
@slrelease slrelease released this 16 Jun 12:03
· 139 commits to main since this release
v0.6.7
8ca2cdd

Features

  • Added a 32K-token serving profile for Qwen3.6-27B on RTX PRO 6000. The base model configuration retains its 4K context window.

Bug fixes

  • Grammar-constrained Qwen3.5-4B requests use a non-speculative profile so structured-output constraints are enforced.
  • Long model startups now use consistent readiness timeouts across the gateway, workers, and Helm configuration. Reapplying unchanged model configuration no longer needlessly unloads models.

Performance improvements

  • LightOnOCR processes multiple pages in bounded batches, preserving page order and handling different image sizes.
Assets 2
Loading