# Release Notes: Multi-level Video Understanding

## Version 2026.2

<!--date TBD-->

This release moves the pipeline to **on-device AI services** and adds runtime-extensible prompt tasks and subtitle-aware summarization.

**New:**

- **On-device AI Services**: vLLM-IPEX-serving and multilevel-video-understanding run locally on a single Intel® PTL 358H host with shared system RAM — no discrete accelerator or remote inference cluster. A single `Qwen3.5-35B-A3B` model serves both the VLM (captioning) and LLM (aggregation) roles from one OpenAI-compatible endpoint.
- **Dynamic prompt tasks**: register/update/delete custom summarization tasks at runtime via the `/v1/tasks` API. Two built-in tasks ship by default — `summary` (English) and `summary_zh` (Chinese); everything else is user-registered.
- **Subtitle-aware summarization**: pass subtitles alongside a video for extra per-chunk context, or run **caption-only** mode (`video: "none"` + subtitles) to summarize a pre-built event log without any VLM inference.
- **Single env file**: all deployment variables consolidated in `docker/set_env.sh`, with a default 64 GB / 32 GB-swap on-device profile.

**HW used for validation**:

- Intel® Core™ Ultra (Panther Lake, PTL 358H) with integrated GPU sharing system RAM
- 64 GB system RAM + 32 GB swap

## Previous Releases

- [Release Notes 2025](./release-notes/release-notes-2025.md)


:::{toctree}
:hidden:

Release Notes 2025 <release-notes/release-notes-2025.md>

:::