Skip to content

v0.8.16

Latest

Choose a tag to compare

@github-actions github-actions released this 03 Sep 21:56
· 4 commits to master since this release
5d400cf

Provider v0.8.16 (Swift CLI)

Binary hash: f5f2242d53fbc7d96e03e4cf49877dbf51f4c5a9c28baae59055070033089ab3
Bundle hash: 3847c009d635f126fa03b10e2bb7a6c08d28ec96cc3ae6d7180a9c441132cca9
Metallib hash: 4ffbbac48a99b495916c3fa0921ce813eb554de1a09aff408d9b5a5a8053e6b0 (built from mlx 734241bbff26)
Signed by: Developer ID Application: Eigen Labs, Inc. (SLDQ2GJ6TL)
Notarized: yes
Min macOS: 14.0

Install

curl -fsSL https://api.darkbloom.dev/install.sh | bash

What's Changed

  • fix(registry): keep Qwen3-VL vision off M5 by @Gajesh2007 in #781
  • feat(routing): add model-specific first-content deadlines by @Gajesh2007 in #787
  • feat(provider): MTP policy lives in the weights — auto activates embedded Qwen heads only by @Gajesh2007 in #777
  • chore: adopt MLX core 0.32.2 across the inference stack by @Gajesh2007 in #790
  • coordinator: survive fleet upgrades without trust collapse by @Gajesh2007 in #778
  • fix(console-ui): keep map tooltips visible and add Qwen branding by @prasiddhnaik in #624
  • feat(routing): fresh-state capacity quotes, dispatch plans, governed hedging by @Gajesh2007 in #792
  • fix(registry): stop serializing full-fleet provider scans by @Gajesh2007 in #793
  • fix(auth): accept hashless provider registrations by @Gajesh2007 in #796
  • fix(auth): retain hashes across trust reuse by @Gajesh2007 in #797
  • fix(auth): make release-policy gate provable by the real fleet by @Gajesh2007 in #798
  • Add DogStatsD histograms for X-Timing latency decomposition by @hankbobtheresearchoor in #723
  • fix(routing): harden dispatch against retry-storm CPU collapse by @Gajesh2007 in #799
  • fix(provider,coordinator): per-model activation floors — measured, hardened, rebased for v0.8.16 (supersedes #654, addresses #653) by @Gajesh2007 in #791
  • test(provider): gate the destructive-epoch capability test on DEBUG so release test builds compile by @davidtai in #812
  • feat: system profiler — per-attempt request waterfall, provider/engine profile, fleet snapshots by @Gajesh2007 in #809

New Contributors

Full Changelog: v0.8.15...v0.8.16