Gradio Distills Qwen-Image 2.1's 9B Prompt Rewriter Down to 0.8B for Laptops
Gradio · x · 2026-09-23
Qwen-Image 2.1 draws best when a 9B prompt rewriter expands your request first — 20GB in bf16, a 1,700-word system prompt, and 1,600 thinking tokens. Gradio distilled it to 0.8B that fits on a laptop. Follow-ups: built end-to-end by ML-Intern in HuggingChat; models, the 8,797-request dataset with teacher thinking traces, all eval predictions, and training/eval code are open under the non-commercial Qwen Research License.
Related event: Gradio Open-Sources Distilled Qwen Prompt Rewriter Down to 0.8B(3 posts)→
More from Multimodal
- Multi-path video effect live demo made in minutes with H3 Max Director on fal — OdinLovis · 2026-09-23
- Early GPT-6 Sol access: 1-hour tasks done in 20 minutes with half the tokens — flavioAd · 2026-09-23
- AntLing open-sources 6B Ming-Image-0.1-Design, #1 open-weight model on UI/UX design leaderboard — Sitkin_Marrel · 2026-09-23
- NTU UMM study: generation training boosts understanding in native multimodal models, but naive sharing conflicts — jiqizhixin · 2026-09-23
- Procedural 3D gen game outshines Sandbox and Decentraland's $400M metaverse bet — Daniel_Farinax · 2026-09-23
- AI-made mini-drama scripted by Pocket FM's Sherpa and rendered in Seedance 2.5 goes viral — chrisfleck · 2026-09-23