Skip to content

v2.10.0

Choose a tag to compare

View all tags
@jashansinghTT jashansinghTT released this 21 Aug 20:50
· 2 commits to main since this release
v2.10.0
1d5505b
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

🚀 New Features

  • Apps Marketplace: New Apps page that connects locally deployed chat models to external AI tools — launch an app from the catalog and it comes up already wired to your model, with setup guides for the ones that run outside TT Studio (replaces the old Coding Agents page)
  • Hybrid RAG Retrieval: Reworked retrieval with reranking, structure-aware chunking, and server-side orchestration for noticeably better answers over large document sets
  • RAG & Search in the Voice Agent: The Voice Agent can now answer from a RAG collection and reach the search agent
  • Voice Input via Deployed STT: Chat voice input routes through a deployed speech-to-text model end to end; chat attachments have been removed
  • Deployment Recovery: Deployments whose container is still running are recovered on restart instead of being orphaned
  • Media Download Progress: Media-model weight downloads report live progress that survives a page refresh
  • --purge-model: Uninstall specific models instead of wiping everything — a shared image is only removed once the last model using it is gone
  • --no-clear: Start without clearing the terminal and stream the full startup detail
  • Clearer Launcher Diagnostics: The launcher explains which image source it picked for each service and why, and reports service failures as readable diagnoses
  • Hugging Face Token Guidance: Onboarding now leads with the HF token step, and the UI warns when no token is set and explains deploys stuck at 0%
  • Token-Free Public Models: Public models deploy and download without a Hugging Face token

🤖 Model Support

  • Qwen3.5-9B (Blackhole): Added as an experimental model, with deploy timeouts fixed and workflow logs streaming during bring-up
  • Gemma4-31B-IT: Added to the model catalog
  • Qwen3.8-27B: Added for p300x2
  • Qwen3.6-27B: Added for p300x2, alongside the tt-inference-server artifact bump to v0.20
  • Z-Image-Turbo: Added for p300x2
  • Catalog Cleanup: Removed gemma-1.1-2b-it

🐛 Bug Fixes

  • Fixed the AnythingLLM container failing to start
  • Fixed app launch issues in the Apps marketplace
  • Fixed RAG upload failures for .log files, double uploads on drag-and-drop, and broken empty collections left behind by failed uploads
  • Fixed a stale model dropdown after a board reset, the warmup clock resetting on page refresh, and the hardcoded video-generation ETA
  • Fresh installs now hand the agent a model token without hand-editing .env
  • Hid auto-managed fields from the settings dialog, trimmed the welcome wizard, and added a settings shortcut
  • Host-service cleanup no longer hides the sudo password prompt behind a spinner

🔧 Technical Improvements

  • Model weights download into the Hugging Face cache instead of named Docker volumes, so re-deploys reuse what's already on disk
  • Container images are published to GHCR and pulled by run.py instead of built locally (--build-images still builds)
  • Removed the CUDA dependency chain from the backend image, cutting it from ~11GB to ~4GB
  • Added a per-model dev-catalog override for deploys
  • CI enforces conventional commit titles on PRs, with release PR titles allowed through
  • Added SECURITY.md with a vulnerability-reporting policy

Full changelog → #1247

Assets 2
Loading