36M-param open-source MCP server lets LLMs watch videos locally on a laptop CPU
Safe-Escape197 · reddit · 2026-11-21
A Reddit developer released v1.0 of tiny-vision (Apache-2.0), an open-source tool that lets LLMs process video entirely locally via an MCP server with summarizevideo, describevideo, and captionimage tools — tested in Claude Desktop and Claude Code, fully offline with LM Studio.
How it works
- Frames are picked at scene cuts (hard cuts + camera drift), not fixed intervals, so timestamps land on real edits
- Each shot is captioned at start/middle/end, then similar shots are grouped into scenes
- 36M params total: frozen DINOv2-small encoder + a 14M adapter/caption decoder trained from scratch on COCO
Numbers (Apple M5, COCO test): caption quality 89% of florence-2-base (CIDEr 1.196 vs 1.348) while being 6x smaller, 8x faster on GPU and 75x faster on CPU (54ms vs 4s per image). A CLI mode pipes plain text into any LLM. Limitations: no audio, no motion understanding, no OCR/counting.
More from coding & agent
- Windows Terminal's agentic "Intelligent Terminal" becomes a daily driver — dfinke · 2026-11-21
- If it's verifiable, it can be one-shotted: an engineering workflow for AI-built systems — imdigitalashish · 2026-11-21
- Have Claude spin up a custom wandb-style training dashboard in minutes — rishabh16_ · 2026-11-21
- Jev is just a classifier — but it replaces months of ML team work in minutes — viksit · 2026-11-21
- omarsar: Stop automating everything — help your coding agents write better code instead — omarsar0 · 2026-11-21
- Dev launches remote MCP server for live Pinnacle sports odds, queryable from Claude Desktop — Significant-Sun-9735 · 2026-11-21