v0.7.0

[0.7.0] - 2026-09-16

Added

• ltx2: added the ltx-2.5-distilled model (Lightricks/LTX-2.5) and made it
  the default. distilled (LTX-2) remains available. LTX-2.5 adds the request
  options auto_duration, enhance_prompt, pipeline="dfr", and
  video_decoder="diffusion". Run kiapi activate --family ltx2 to install the
  new mlx-video and the gated LTX-2.5 weights (about 83 GB including the prompt
  enhancer and DFR adapter).
• setup: HfSnapshotResource accepts allow_patterns so a model can download
  only the files it loads from a large repo.
• chat: added the qwen3.8-27b model (mlx-community/Qwen3.8-27B-4bit, aliases
  qwen3.8, qwen3_5, qwen3-vl, vlm). The last three moved from
  qwen3.6-27b, which keeps only qwen3.6. It runs on the existing qwen3_5
  handler (text + image, Hermes/XML tool calls, reasoning off by default).

Fixed

• chat: image and video input failed on every model with
  TypeError: repeat(): incompatible function arguments since mlx 0.32. Fixed by
  updating mlx-vlm (below).
• chat: Qwen3-Omni now receives its deepstack visual features. mlx-vlm 0.6.3
  computed them but dropped them before the decoder, so image/video input used
  less of the vision tower than intended.
• chat: Qwen3-Omni video input no longer decodes garbage or crashes the server
  with a Metal GPU address fault on mlx-vlm 0.7.1. The deepstack mask and
  features are now windowed to each prefill chunk (patch H, upstream #2099).
• chat: Qwen3-Omni placed the video deepstack rows at the wrong positions when a
  prompt had both an image and a video. Patch C now rewrites that join as in
  upstream PR #2257 instead of grafting mx.where / mx.scatter onto mlx.

Changed

• chat: updated mlx-vlm from 0.6.3 to 0.7.1 (still pinned exactly). The
  streaming UTF-8 and mx.repeat patches are removed because upstream fixed
  both; the stereo-audio, mx.where/mx.scatter, and stream-text patches stay.
  mlx-lm is now declared directly because mlx-embeddings' Qwen3-VL model
  imports it and mlx-vlm no longer pulls it in.
• Restructured the repository from a uv workspace (packages/kiapi) into a
  single package: the source now lives in src/kiapi/, the tests in tests/,
  and the release workflow builds and publishes kiapi directly. The package
  README and CHANGELOG are merged into the root ones, and the root VERSION
  file is replaced by the version in pyproject.toml.
• Updated dependencies. FastAPI moves to >=0.141 (build_openapi now walks routing.iter_route_contexts to handle the lazy included routers of FastAPI 0.137+; the generated OpenAPI documents are unchanged), and the numpy<2.5 cap is lifted (numba >= 0.67 supports numpy 2.5; the out-of-band LTX-2 install needs numba>=0.67).
• Refreshed the locked dependencies: torch 2.14, torchvision 0.29,
  huggingface-hub 1.30, tokenizers 0.23.2, anyio 4.15, and ruff 0.16.6.
• Updated the GitHub Actions used by CI, the PyPI release, and the Pages deploy
  to their current majors. upload-pages-artifact now sets
  include-hidden-files: true so public/.nojekyll keeps being published.
• Raised the mise pinned in CI and the release workflow from 2026.5.0 to 2026.9.1,
  the version mise run ci is verified against locally.
• Replaced the Dependabot configuration: the npm ecosystem entry pointed at the
  package.json removed in 0.6.0, so it is dropped in favour of the uv and
  github-actions ecosystems. mlx-vlm is ignored there for the same reason it
  is pinned.

Removed

• Removed the Node tooling (package.json / pnpm / firebase-tools) that existed only for the retired GCP relay setup task. This clears all open Dependabot alerts, which were transitive dependencies of firebase-tools.