Test shows Qwen3.8-27B excels at video understanding with precise timestamping
vanstriendaniel · x · 2026-08-19
A test reveals that Qwen3.8-27B performs strongly on video description tasks. In a single request, the model processed an 11-minute film from 1935, identifying 96 timestamped events and quoting on-screen text verbatim. Verification confirmed timestamp accuracy within approximately 2 seconds across the entire film. The task ran on a single GPU via Hugging Face Jobs, with a WIP recipe shared by the author. Another user adapted this workflow using vLLM and H-100 GPUs for bulk processing of TV archive commercials, reporting excellent results.
More from Apps
- Trinka AI Launches Confidential Data Plan: Zero Storage or Training — Faheem_uh · 2026-08-19
- Siraj Raval shares his $90 full-stack AI film production toolkit — sirajraval · 2026-08-19
- "Inspiration Prompter" Adds 2K and 4K Export Support — yihui_indie · 2026-08-19
- Same Prompt, Three Tools: Claude vs ChatGPT vs Figma Make UI Generation Face-Off — Tegadesigns · 2026-08-19
- Extracting nested financial tables with Sparrow & OCR — andrejusb · 2026-08-19
- Apps need a design laboratory: tweak values and use natural language for vibes — threepointone · 2026-08-19