Skip to content

v0.10.14

Choose a tag to compare

@github-actions github-actions released this 17 Nov 05:47
· 7 commits to master since this release
v0.10.14
d8af266

Darkbloom provider 0.10.14 plays more videos and images, keeps tool calls and JSON answers in shape, brings Gemma 4 26B
back to two-Mac pairs, and retires MDM: Macs are verified with App Attest only. Update with darkbloom update.

Highlights

Models and quality

  • Qwen 3.6 35B ends a thought that keeps repeating itself with an answer, instead of running out of tokens, as Qwen 3.5
    already does.
  • Qwen models read a video the way their reference implementation does: up to 32 frames, each pair of frames labelled
    with its time in the clip. They answer more questions about a video correctly; the first token can take a few
    seconds longer.

Images and video

  • MP4 and QuickTime videos sent inline are played whatever type they are labelled with. Other video formats get a
    clear "unsupported" answer instead of "invalid".
  • Qwen models take up to 64 images in one request and Gemma 4 26B up to 24, for example a video's frames.
  • GLM-5.3-Flash plays long videos instead of refusing them (their frames are made smaller to fit), and sets aside memory
    for images and videos by their real size, so it turns fewer of them away.
  • A very short 4K clip that GLM-5.3-Flash can't take gets a clear answer at once ("lower the video's resolution")
    instead of being retried on other Macs.

Tool calls and JSON

  • A tool call that leaves out an argument, or adds one the tool doesn't declare, comes back as a tool call on every
    model, instead of tool markup in the text or an error.
  • GLM-5.3-Flash's JSON answers follow the requested schema by default. They take a little longer to generate (about
    0.7–0.85× the speed); the time to the first token is unchanged.
  • Tool definitions that refer to shared parts of a schema ($ref) reach the model in full on your local API.

Two-Mac pairs

  • Two-Mac pairs serve Gemma 4 26B again, on its current build, split by layers or across both Macs at once.
  • A pair that serves only you comes back by itself after darkbloom update on both Macs.

Security and MDM retirement

  • Darkbloom no longer uses MDM. Macs are verified with App Attest only. darkbloom unenroll names the old Darkbloom
    profile and opens System Settings so you can remove it; Darkbloom no longer asks for an admin password to read
    profiles.
  • Stricter request limits on the provider's local API and on the Darkbloom website's API.
  • Tighter memory accounting when a Nemotron 3.5 Super video request is cancelled.
  • The security stack is intact: end-to-end encryption, attestation, signing and weight hashes are unchanged.

Command line

  • darkbloom update says when the disk is too full for an update and how much space it needs.
  • On a pair that serves only you, darkbloom update says it asked for a restart without waiting for the other Mac,
    instead of promising one "now".

Upgrade notes and known issues

  • Pairs: both Macs of a pair must run the same version. A pair on 0.10.13 updates itself.
    darkbloom update --rollback puts back 0.10.13.
  • Known issue, pairs that serve only you: the update to 0.10.14 still works the 0.10.13 way: it waits, and the
    pair usually comes back by itself within minutes. If it stays down, run darkbloom cluster stop --local-only --service,
    then darkbloom start, on each Mac. From 0.10.14 on, darkbloom update on both Macs is enough.
  • The old MDM profile: if your Mac still has the "Darkbloom Provider Enrollment" profile, darkbloom unenroll shows
    it and opens System Settings to remove it. Nothing else is needed; serving doesn't depend on it.
  • Known limitation (unchanged since 0.10.0): Nemotron 3.5 Super is served only by the background provider;
    darkbloom start --local --here skips it.

For operators: turning a change off

Set the variable before darkbloom start (or darkbloom cluster); the installed service keeps it.

Variable What it restores
DARKBLOOM_GLM_JSON_ANSWERS=0 GLM-5.3-Flash JSON answers as in 0.10.13 (steered by the prompt, then checked)
DARKBLOOM_LOOP_EXIT_THINKING set to a list without qwen3_5_moe Qwen 3.6 35B (and 3.5 35B) thought loops run to the token limit
DARKBLOOM_QWEN_VIDEO_MAX_FRAMES=8 and DARKBLOOM_QWEN_VIDEO_MAX_TOKENS=1024 Qwen's earlier video size
DARKBLOOM_MAX_IMAGES_PER_REQUEST=16 One limit of 16 images a request
DARKBLOOM_GRAMMAR_ONLY_TOOL_PARSING=0 0.10.13's tool-call parsing
DARKBLOOM_GLM_MEDIA_GEOMETRY_RESERVATION=0 GLM-5.3-Flash's earlier media memory charge
DARKBLOOM_GLM_LONG_VIDEO_FIT=0 GLM-5.3-Flash refuses long videos again
DARKBLOOM_CLUSTER_LOCAL_UPDATE_RESTART=0 0.10.13's update of a pair that serves only you
DARKBLOOM_TOOLSCHEMA_REF_RESOLVE_MODELS=off Local API tool schemas with $ref passed through as before

The full technical changelog is in CHANGELOG.md.

Provider v0.10.14 (Swift CLI)

Source: d8af266
Build and SDK qualification: https://github.com/Layr-Labs/d-inference/actions/runs/44021655814
Signing: https://github.com/Layr-Labs/d-inference/actions/runs/44021655814
Binary SHA-256: 6edb1ce351bd132c24486e9a66b1e27e47ec60b6e099bc529b022c456383cfc6
CodeDirectory SHA-256: c394bb5574f47d24858ec8894eab979958d122c26d8c4298da42f11b851e5aeb
Bundle SHA-256: 3e700afa30fd8a3da99981d53208534c4131af7b192bfad0e3e50aa22348cb80
Metallib SHA-256: 5b97893cc20e6e9651edf3284d970488af9a9bdd64820231e95fee760eec70ac
Signed and notarized: yes
macOS deployment target: 14.0 (build, tests and signed runtime smoke: macOS 27)

Install

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

Full Changelog: v0.10.13...v0.10.14