Fine-tuned Qwen3.5-4B builds prettier working web pages on an 8GB laptop GPU
MuzafferMahi · reddit · 2026-11-21
A developer released Qwen3.5-4B-Web-Designer (Apache-2.0), a QLoRA fine-tune that turns plain requests into self-contained HTML pages. On 74 held-out prompts browser-checked for JS errors, dead buttons and failed loads, working pages rose from 35% to 53% and completed pages from 75% to 95% vs base.
- Training: 1,443 prompt-to-page examples from LMArena WebDev prompts + own requests, answered by DeepSeek V4.1 Flash with a short design plan; every page browser-verified before entering training
- QLoRA rank 16, 2 epochs, on a single RTX 4060 Laptop (8 GB)
- Ships as GGUF: one llama-server command, temp 0.7 / top-p 0.95 / top-k 20
- 148-page public gallery including failures; independently coincides with ccore's Web-Coder
More from coding & agent
- Steve Yegge: your agent remembers every instruction, including ones you reversed — Steve_Yegge · 2026-11-21
- Coding Is Dead, Not Software: Good Software Is a Set of Strong Opinions — pramodk73 · 2026-11-21
- Juan Benet: AI agent offered to rebuild DocSend feature it lacks — SaaS must get moldable — juanbenet · 2026-11-21
- Tips for building voice agents with the gpt-live-1.1 API: backend and reasoning level matter — pbbakkum · 2026-11-21
- Redditor wires Claude Code to a spare iPhone for 24/7 real-device agents — MasterReading8368 · 2026-11-21
- Cerebras runs Kimi K3.1 Pro at up to 2,000 tokens/sec, hosts $5,000 hackathon — film_girl · 2026-11-21