v0.3.3

0.3.3 (2026-10-13)

Bug fixes

• media: record the clips at 2x for real (8595124)

Documentation

• media: encode every clip's MP4 at 1920 x 1200 (92440b0)
• media: open and close the tour on the logo, at 1920 x 1200 (71e80d1)
• media: open and close the tour on the logo, at 1920 x 1200 (9770722)
• media: re-record the README clips at 1920 x 1200 (ddaa837)
• media: re-record the README clips at 2x (d0d3579)
• media: re-record the README clips at a real 2x (442ff9d)

v0.3.2

0.3.2 (2026-10-13)

Documentation

• media: re-record the README clips on a fresh install (88443e8)
• media: refresh the README clips and add screenshots (3fb33e8)
• readme: show screenshots of translation, songs, presets, workflows, phone, monitor and setup (25bfcc6)
• validation: record a fresh install on a DGX Spark (ea5e581)

v0.3.1

0.3.1 (2026-10-13)

Bug fixes

• serve user files safely, validate inputs, surface lost connections (d4624c6)
• server: serve user files inert, validate inputs and keep the worker alive (2f4f645)
• ui: trap dialog focus, show lost connections and use only tokens (df749e2)

Documentation

• document the environment and add troubleshooting (bd30766)
• readme: drop the line on where the README layout came from (04b8239)
• readme: drop the line on where the README layout came from (29d8dfc)

v0.3.0

0.3.0 (2026-10-13)

⚠ BREAKING CHANGES

• SPARKLINGKIT_* environment variables are no longer read, the containers and images are renamed, scripts/start-sparklingkit.sh is removed, and preferences kept in the browser reset once.
• models, endpoints and API keys can no longer be chosen. Settings saves ignore the services, the *_MODEL and *_API_KEY variables are ignored, the removed start options refuse other values, and workspaces configured by hand go through setup again.

Features

• guard heavy jobs and start-up against running out of memory (9a5e8d3)
• llm: make Saluki 27B the default language model (bd5145e)
• make Sparkfire a fixed model stack with Saluki (7571cb2)

Bug fixes

• dgx: describe the fixed stack in start-sparkfire.sh's messages (9bfedc2)
• dgx: stop pointing the memory check at spark-switch.sh (d8b3f5f)
• media: keep the demo copy's own monitor and SearXNG (e8befcc)
• mindmap: ask for mind maps with thinking off (0a83557)

Documentation

• add a contributing guide (39ef825)
• bring the docs up to date with the fixed stack (eac44fe)

Code refactoring

• rename everything from sparklingkit to sparkfire (80a961b)

v0.2.0

0.2.0 (2026-10-12)

Features

• dgx: start the music service with --music-backend yue2 (6e12783)
• music: add a YuE2 music generation service (16e96b0)
• music: make songs from lyrics and a style (b8c01c6)
• queue: run heavy GPU jobs one at a time through a shared lane (15da951)
• ui: give Sparkfire a warm palette and an ember accent (5d3a758)
• ui: give Sparkfire its own ember look (a5c5894)
• ui: mark the active page and running work in ember (972bf08)
• ui: switch to Geist and Bricolage Grotesque (d0ae715)
• ui: tighten corners and drop pill buttons (e549e8e)

Bug fixes

• queue: let waiting heavy runs reach the lane (09189ad)
• store: serialize changes to a job's manifest (894bcee)
• ui: give the song style field three lines on phones (db11c03)
• ui: show why a queued job waits in Running now (51cd81d)
• ui: take graph and preview colours from the theme (1b7fbea)

Documentation

• describe the ember interface (98ae42c)
• document music generation with YuE2 (5ca2103)
• media: record a second video waiting for the GPU (7f31b2a)
• media: record the tour and clips with the ember interface (a6689ab)
• readme: explain that heavy jobs take turns (27ba622)
• validation: measure what each service costs a song and the reverse (23ac59f)

v0.1.0

0.1.0 (2026-10-11)

⚠ BREAKING CHANGES

• the published image moves from ghcr.io//sparklingkit to ghcr.io//sparkfire, and distribution/install.sh is gone.
• dgx: start-dgx-spark.sh no longer builds, downloads or starts the video service unless --video-backend fasth3 is passed, and a restart without it stops a running video service.

Features

• asr: add a Parakeet adapter with the OpenAI transcription route (d7553e3)
• asr: build sentences, subtitle cues and paragraphs from word timestamps (151970d)
• asr: make Parakeet the default speech recognition backend (a90ab75)
• asr: package parakeet.cpp for the DGX Spark (611b327)
• asr: request word timestamps from Parakeet (7b56582)
• asr: write sentence lines and subtitle cues from word timestamps (23ff788)
• chat: let the model search the web and read pages (85eabd0)
• dgx: add an opt-in SearXNG for chat web search (b9ee3ae)
• dgx: drop the bundled LLM, Qwen3-ASR and grounding services (79bad49)
• dgx: drop the bundled LocateAnything grounding service (6e3a8c9)
• dgx: drop the bundled Qwen3.6 LLM from the reference stack (51c8ded)
• dgx: drop the Qwen3-ASR backend; Parakeet is the only speech service (4bb1bdf)
• dgx: make FastH3 video generation opt-in (f41e66a)
• dgx: make Parakeet the default speech backend of the start script (c52e28e)
• dgx: run FastH3 video generation in the reference stack (03ff048)
• read SPARKFIRE_* environment variables (a4ffe6f)
• ui: add the Sparkfire mark (d2e3f3a)
• ui: add the Sparkfire mark (3423884)
• ui: hide grounding until a grounding endpoint is configured (662f0c2)
• ui: hide grounding until a grounding endpoint is configured (145980e)
• ui: offer video generation on the workbench instead of grounding (f3372a6)
• ui: rename the product to Sparkfire (dac7baa)
• video: add a FastH3 video service driven through ComfyUI (5c50f1f)
• video: add FastH3 text-to-video with sound (c063bd8)
• video: add the text-to-video module (8b7bb24)
• video: offer 10-second videos (6430549)
• video: offer 10-second videos (d8932db)

Bug fixes

• asr: drop Parakeet's <unk> marks from words (ad960b5)
• asr: keep chunk text when an ASR returns no word times (46888aa)
• asr: keep subtitle cues within their caps when punctuation tokens follow (22012fa)
• asr: run the Parakeet container without root and with a health check (b6baee3)
• asr: send Parakeet chunks of at most 30 seconds (3d5c35f)
• asr: strip Qwen3-ASR language tags anywhere in a transcript (bc69103)
• chat: keep answers going when the page leaves the chat (887cfc4)
• chat: keep answers going when the page leaves the chat (e776c88)
• deps: patch security advisories in production dependencies (980c5c2)
• dgx: default split installs and the demo to the Parakeet model id (2941be1)
• dgx: name the Parakeet engine from its command line (fd0d639)
• dgx: report only the services the stack runs in status (03e95a3)
• dgx: start Parakeet through its readiness variables (f0c6a61)
• dgx: stop and remove containers of the retired services (2445239)
• media: forward only the status route through the demo proxy (e257921)
• patch dependency advisories, add CI/CD and make the fork public-ready (cfd3145)
• ui: point footer, update check and onboarding at the fork (403ac9a)
• video: create ComfyUI's base folders before starting it (12e4999)

Dependencies

• dgx: bump python-multipart to 0.0.31 in the adapters (eed61c5)

Documentation

• add per-container memory tables and Saluki's context size (f095164)
• add the Sparkfire notice above SparklingKit's (00bb9c8)
• advise keeping the queue at two workers on a Spark (acfe2d6)
• agents: fix the memory rule after the stack lost its LLM (396b088)
• asr: design Parakeet as the default speech recognition backend (531cd6e)
• asr: document Parakeet as the default speech backend (5790d88)
• asr: plan the Parakeet backend and refine its timing rules (43256b3)
• describe chat web search (89ecd7d)
• describe FastH3 video as on by default (37bdeed)
• describe Sparkfire and credit SparklingKit (cb9aff0)
• describe the CI and release process and Conventional Commits (1dd91c1)
• dgx: explain upgrading from the six-model stack (373d980)
• media: record the tour with video and web search at full resolution ([8ce6928](8ce69287e65...