Gradio distills Qwen's 9B prompt rewriter into an 812MB 0.8B model that fits a laptop
Gradio · x · 2026-09-23
Gradio released Qwen-Image-2.1-PE-T2I-Pocket, distilling Qwen-Image 2.1's official 9B prompt rewriter into 2B and 0.8B student models.
- The original 9B rewriter runs 20GB in bf16, uses a 1,700-word system prompt, and thinks for 1,600 tokens before answering. The 0.8B student needs no system prompt and no thinking — it goes straight to the JSON the image model wants in 450 tokens (28% of the teacher's), and ships as an 812MB GGUF that fits on a laptop.
- The whole pipeline was run end to end by an ML-Intern agent in HuggingChat: it wrote 8,797 requests in 18 languages, labeled them with the 9B teacher, trained both students with TRL, evaluated on text and rendered images, and pushed everything to the Hub. Humans only approved the budget and read the report.
Weights are open on Hugging Face under ML-Intern-lab.
Related event: Gradio Open-Sources Distilled Qwen Prompt Rewriter Down to 0.8B(3 posts)→
More from coding & agent
- MatBrain splits reasoning from tool use: two models screen 30,000 crystal candidates in 48 hours — bravo_abad · 2026-09-23
- Firecrawl Raises $75M Series B, Launches Alexandria Knowledge Library for AI Agents — omarsar0 · 2026-09-23
- Five AI agents bypassed a permissions broker in ten minutes using 'start' instead of 'stop' — TrifleHopeful5418 · 2026-09-23
- Shopify CEO who pushed staff to use AI now 'horrified' — the 'Slop Grenades' story — srchvrs · 2026-09-23
- Lovable adds Claude Opus 5.5 and GPT-6 Sol, auto-routing between frontier models — AlexandrePesant · 2026-09-23
- Dev laments agents built around KV caches, wants inference-first chips — dbreunig · 2026-09-23