v0.8.0

[0.8.0] - 2026-09-26

Dependencies

• mflux 0.19.1 -> 0.20.0, pinned to a kiarina fork commit (144a6bec) that adds
  mflux-community/mflux#741 on top of 0.20.0. 0.20.0 also fixes the seedvr2
  mx.repeat failure.

Removed

• chat: removed the qwen3.6-27b model (alias qwen3.6). qwen3.8-27b uses the
  same handler, modalities and memory footprint; vlm selects qwen3.8-flash-next.
  Remove the downloaded weights with kiapi deactivate --repo mlx-community/Qwen3.6-27B-4bit.

Changed

• chat: model is now required on /v1/chat/completions; there is no default
  chat model. Requests without it return HTTP 422. The chat models differ in
  accepted input modalities (only qwen3-omni takes audio/video) and memory
  footprint (qwen3.8-flash-next keeps about 74 GiB resident), so an implicit
  default could silently evict other models or reject media.
• chat: the aliases vlm and qwen3.8 now select qwen3.8-flash-next
  (previously qwen3.8-27b). qwen3.8-flash-next also answers to
  qwen3.8-flash and flash-next; its qwen4_exp alias was removed.
  qwen3.8-27b keeps qwen3_5 and qwen3-vl.
• BREAKING: chat: removed the max_tokens_cap setting
  (KIAPI_CHAT_MAX_TOKENS_CAP, 4096). max_completion_tokens is no longer
  capped by the server; generation stops at max_completion_tokens or when the
  prompt plus the output fills the model's context window, whichever comes first.
• chat: the default max_completion_tokens is now 1024 (was 512).
• chat: non-streaming requests now run through stream_generate like streaming
  ones, so the context window bound applies to both.

Added

• Web UI at /: a dashboard (server, queue, memory, setup), model setup
  status with copyable kiapi activate commands, jobs, files with previews, and a
  guide and API reference for every family, in light and dark themes. Every
  family can be run from its Playground: forms are built from the family's
  OpenAPI schema, generation runs as a job with live progress, inputs come from
  uploads or stored files, and "Write with chat" drafts prompts with a chat model
  that reads the family guide. Chat has its own conversation view with every
  request parameter (tools, tool choice, parallel tool calls, token limits,
  sampling, template kwargs, streaming on or off), tool-call results, and token
  usage. A "?" beside each field shows its full description, and a floating
  assistant answers questions about the current family from its OpenAPI document
  and, when asked, fills in the form through tool calls (it never submits).

• GET /v1/setup: every model with its setup state, like kiapi status.

• qwen: added the image-2.1 model (Qwen/Qwen-Image-2.1, aliases
  qwen-image-2.1, qwen-2.1). One resident model serves both /generate
  (text-to-image) and /edit (up to 10 reference images), outputs RGBA, and
  renders up to 2752 px. Defaults are 40 steps, guidance 1.0 and q8. It does not
  take init_image or loras. Its weights are under the non-commercial Qwen
  Research License. Editing comes from the pinned mflux fork (144a6bec,
  mflux-community/mflux#741); installs with official mflux do not register it.

• qwen: jpeg output flattens transparent areas onto white.

• chat: Qwen3.8-Flash-Next reuses unchanged history when images are appended,
  like Qwen3.8-27B, through the pinned mlx-vlm fork (6581ba8c).

• chat: added the qwen3.8-flash-next model (mlx-community/Qwen3.8-Flash-Next-4bit,
  aliases qwen3.8-flash, flash-next, qwen4_exp). It loads through a view that
  memory-maps its n-gram (PLE) table, which keeps about 80 GB resident instead of
  111.5 GB, and raises the open-file limit the mapped table needs.

• chat: Omni reuses unchanged prefixes when images, audio clips or videos are
  appended, including demuxed audiovisual inputs, through a pinned mlx-vlm fork.

• chat: the pinned engine supports multiple audio clips with independent feature
  extraction/encoding and corrected CNN lengths and chunk masks.

• chat: Qwen3.8 reuses unchanged image history when new images are appended,
  using a pinned mlx-vlm fork in uv-managed checkouts. Official-engine installs
  retain the conservative whole-request fallback.

• chat: client disconnects now cancel queued work or stop running generation at
  the next token boundary. Cancellation uses the existing canceled job state
  and safely clears APC state when a generator closes early.

• chat: responses report APC reuse through the OpenAI-compatible
  usage.prompt_tokens_details.cached_tokens field. Streaming requests support
  stream_options.include_usage and emit usage before [DONE] when requested.

• chat: bounded, memory-only automatic prefix caching for Qwen3.6 / Qwen3.8
  text and images, and Qwen3-Omni text, image, audio, video, and image + video.
  Media content hashes and video options protect cache identity, and Omni
  restores complete-prompt positions when reusing media prefixes.

• chat: GET /v1/models returns each model's context_window, read from the
  model's config.json (null until the model is set up).

Fixed

• chat: suppress extra streamed tool names when parallel_tool_calls=false.

• memory: include idle chat caches in cross-model eviction and transient reservations,
  and derive active cache headroom from the configured APC capacity.

• chat: finish_reason is now "length" when generation stops at
  max_completion_tokens or the context window. It was always "stop".

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.

v0.6.0

[0.6.0] - 2026-09-01

Removed

• BREAKING: Removed the relay transport and retired the kiapi-relay and kiapi-proxy packages. kiapi run --relay, the relay-gcp extra, the relay field of /health, and the KIAPI_RELAY_* settings are gone. To reach kiapi from other machines, expose it over your own private network layer instead — for example, run tailscale serve --bg --https=8500 8500 on the kiapi machine and point clients at https://<machine>.<tailnet>.ts.net:8500. The published kiapi-relay / kiapi-proxy distributions remain on PyPI as-is but will receive no further updates. (The unreleased GCP RTDB watch read-timeout fix was removed together with the relay.)

v0.5.3

[0.5.3] - 2026-07-30

Fixed

• kiapi: kiapi service install now preserves PATH in the launchd property list so background capabilities can find external tools such as Homebrew FFmpeg.

v0.5.2

[0.5.2] - 2026-07-29

Fixed

• Publish each workspace package with its own PyPI Trusted Publishing token so a release can upload multiple projects.

v0.5.1

[0.5.1] - 2026-07-29

Added

• kiapi: kiapi run --relay none explicitly disables the relay, overriding a relay enabled in user settings.
• kiapi-relay: KIAPI_RELAY_DEFAULT=none (and other KIAPI_RELAY_* env vars set to none) is now parsed as unset, so the relay can be disabled via the environment.

Fixed

• kiapi-relay: Reconnect the GCP RTDB request stream when Firebase reports an expired credential or canceled subscription, instead of leaving the relay unable to receive new requests.

v0.5.0

[0.5.0] - 2026-07-11

Fixed

• kiapi / kiapi-proxy: service install now also pins XDG_CACHE_HOME in the launchd property list so the background service uses the same cache directory and single-instance lock as the interactive CLI.

v0.4.0

[0.4.0] - 2026-07-11

Added

• kiapi: Added kiapi service show to print the installed launchd property list.

Changed

• kiapi-relay: The gcp:setup task no longer passes --scopes to
  gcloud auth application-default login. The default ADC scopes already
  include cloud-platform, which covers the GCS and RTDB access the relay
  needs, so the extra scopes were unnecessary.
• kiapi-relay: The gcp:setup Impersonation method now runs
  gcloud auth application-default login itself instead of only reminding the
  user to, since ADC is the base credential the impersonation chain mints SA
  tokens from, and grants roles/iam.serviceAccountTokenCreator to the actual
  ADC principal rather than the active gcloud CLI account.

Fixed

• kiapi: kiapi service install now pins the current XDG_CONFIG_HOME and XDG_DATA_HOME values in the launchd property list so the background service resolves the same user settings and data directories as the interactive CLI.
• kiapi: The hot-reload worker subprocess (kiapi run --debug) now loads the user settings file in the ASGI factory, so relays configured only in user settings (for example the GCP relay's database_url/bucket) start correctly. Previously kiapi run --relay gcp --debug failed at startup with "required field is not set" because the reload subprocess never ran load_user_settings().

v0.3.0

[0.3.0] - 2026-07-02

Added

• Manage Node dev tooling with pnpm through a root package.json; firebase-tools is now a project-local dev dependency installed by mise run setup (pnpm install) rather than a global npm install -g, and mise puts it on PATH via node_modules/.bin.
• kiapi-relay: Added a gcp:setup mise task (run from packages/kiapi-relay/) that interactively provisions the GCS bucket, Firebase Realtime Database instance, and authentication for GCPRelay, then prints the kiapi YAML to paste with kiapi config edit. It uses the project-local firebase-tools. The task verifies firebase-tools has its own login (via firebase login:list) instead of relying on firebase projects:list, which also succeeds through the Application Default Credentials fallback and then fails the Realtime Database calls with a quota-project 403; the RTDB creation failure message now points at firebase-debug.log and lists both the missing-login and Blaze-plan causes. The GCP relay README was rewritten around this task.
• kiapi: GET /health now reports the status of the relay started with the server in a relay field (name, running, failed), or null when no relay is configured.
• kiapi-relay: Added a name attribute to the Relay protocol, populated by the relay registry through a factory_wrapper and shared via a new BaseRelay base class. RelayRunner.status() returns a new RelayHealth view (name, running, failed).
• kiapi: Added a request method to the Relay protocol and implemented it on LocalRelay and GCPRelay, promoting the relay request client from the verification scripts into the relay packages. Responses are returned as RelayResponse, with binary bodies materialized to a temporary file the caller owns.
• kiapi-relay: Relay participants derive a stable node_id from a data directory (get_or_create_node_id), and discover a target node through liveness heartbeats published under liveness/{node_id} as part of the watch lifecycle (heartbeat_interval_s/liveness_ttl_s settings); Relay.request fails with no_relay_node when none is fresh.
• kiapi-proxy: Expanded the CLI to mirror the kiapi command layout so the proxy is managed independently: config (init/show/edit/template) manages a user settings file separate from kiapi's (holding kiapi_proxy.api and kiapi_relay settings, loaded on every command); check --relay local|gcp sends a single request (default /health, overridable with --path) through the relay to a live kiapi node and prints the response without starting the server, so relay connectivity can be verified as a health check (it reuses the persistent relay node_id and holds the single-instance lock like run, failing fast if the proxy server is already running); service (install/start/status/stop/uninstall) manages a launchd user agent (io.github.kiarina.kiapi-proxy) that runs kiapi-proxy run (install pins the XDG_CONFIG_HOME/XDG_DATA_HOME values present at install time into the plist, since launchd does not inherit them, so the service resolves the same config/data directories as the interactive shell and can find the user settings written by config edit).
• kiapi / kiapi-proxy: Each server resolves a persistent relay node_id from its user data directory, injects it into the relay, and acquires a single-instance lock (via kiarina-utils-app, scoped to the user cache directory) so a second process cannot share the same node identity.
• kiapi-relay: Initial release of kiapi-relay, extracted from kiapi. Provides the relay protocol, request/response schemas, the in-process RelayRunner, and the relay request client, plus the local filesystem (kiapi_relay.impl.local) and GCP (kiapi_relay.impl.gcp, available via the gcp extra) relay backends.
• kiapi-proxy: Initial release of kiapi-proxy: a proxy server that forwards incoming HTTP requests to a kiapi instance over a relay (kiapi-relay) and returns the result. Supports JSON responses, file/binary responses, and chat text/event-stream responses re-emitted as Server-Sent Events. Ships a kiapi-proxy CLI and does not depend on kiapi or MLX, so it runs on Linux, Windows, and resource-constrained machines.

Changed

• Reworked mise run verify into a Python driver (scripts/verify.py) that selects a target (--kiapi / --kiapi-relay / --kiapi-proxy, or fzf-interactive), starts and stops the kiapi / kiapi-proxy servers it needs (stopping and restarting the launchd services if they are running), and runs the matching verification scripts. It replaces the old per-capability and relay* task arguments; scope with --family and --relay instead, or run an individual scripts/capabilities/verify_*.py directly. Capability verify output now honours KIAPI_VERIFY_DIR (default .verify), and the driver routes artifacts to .verify/kiapi vs .verify/kiapi-proxy so direct and proxied runs no longer collide. The relay verify scripts were consolidated into scripts/relay/verify_local.py and scripts/relay/verify_gcp.py (which gain the core files/jobs CRUD checks); the capability-heavy scripts/relay/verify.py was removed, since capabilities are now covered through the proxy path. The Makefile verify targets are now verify / verify-fast / verify-kiapi / verify-kiapi-relay / verify-kiapi-proxy.
• Moved development and CI operations from Make recipes into package-aware mise tasks, including the new setup task and namespaced test-assets download task.
• Manage the repository version in a single root VERSION file and release the whole workspace under one shared version. The release pipeline now detects packages with unreleased changelog entries, bumps and publishes only those, and is triggered by a single v<version> tag instead of per-package <package>-v<version> tags.
• kiapi-relay: The relay node_id is now generated automatically and persisted per data directory instead of being configured. The manual node_id/source_node_id settings were removed from the local and GCP backends; clients discover a target node through liveness heartbeats and address responses with their own generated node_id.
• kiapi: Simplified the Relay protocol to a single watch method and moved listener tasks and the HTTP client into the watch lifecycle, removing the explicit close method.
• kiapi: Reworked the relay verification scripts to issue requests through Relay.request via the relay registry factories, removing the duplicated transport client in scripts/relay/_client.py.
• kiapi: Converted the repository into a uv workspace and moved the kiapi package to packages/kiapi/ with a src/ layout. Packaging and lint/test paths are now per-package.
• kiapi: Extracted the relay subsystem into a separate kiapi-relay package. kiapi.core.relay is now kiapi_relay, and kiapi.relay.{local,gcp} are now kiapi_relay.{local,gcp}. The relay-gcp extra now pulls kiapi-relay[gcp].
• kiapi / kiapi-proxy: User-directory resolution and single-instance locking are delegated to the shared kiarina-utils-app package and used directly (kiarina.utils.app) rather than through a core.app re-export layer, removing the private AppSettings/user-directory copy and the direct platformdirs dependency. kiapi's core.app now provides only the AppContext schema; kiapi-proxy's core.app module was removed. Each server sets the app identity by calling kiarina.utils.app.configure(...) at its CLI entry point (and, for kiapi, the ASGI factory used by hot reload). The directory getters return pathlib.Path. For kiapi, the user settings.yaml section for these settings moves from kiapi.core.app to kiarina.utils.app; the override environment variables are KIARINA_UTILS_APP_.

v0.2.0

[0.2.0] - 2026-06-26

Added

• Added an optional plugin-based remote job relay with Firebase Realtime Database notifications, GCS request/response payloads, in-process ASGI dispatch, committed-response recovery, and atomic terminal delivery.
• Added a filesystem-backed LocalRelay for local relay verification without GCP services.
• Added relay verification scripts for LocalRelay, GCPRelay, and end-to-end LocalRelay capability checks.

Changed

• Moved GCP relay dependencies into the optional relay-gcp extra.
• Updated the minimum pydantic-settings-manager dependency to 3.7.0.

Fixed

• Added the OAuth scopes required by the Firebase Realtime Database REST API to GCP relay credentials.
• Added relay support for multipart file uploads such as POST /v1/files.