Qwen 3.8 27B vs Flash-Next vs DeepSeek V4.1 Flash: one-shot landing page build-off
jesdga95 · reddit · 2026-11-18
A single-prompt benchmark had three local models each build a one-file landing page for a fictional NYC photographer on identical tooling (RTX 5090 + 5060 Ti, 128GB RAM).
- Qwen 3.8 27B (NVFP4): 12 min, 81 tool calls (3 failed), 182 tok/s, 3.8MB HTML. Safest design, minor layout flaws, but usable.
- Qwen 3.8 FN (NVFP4): 12 min, 124 calls (4 failed), 303 tok/s, 2.1MB. The author's favorite — mosaic hero reveal, animated captions, live clock, form validation, and self-testing bug fixes.
- DeepSeek V4.1 Flash (IQ1M): 36 min, 155 calls (only 1 failed), 94 tok/s, 1.4MB. Boldest design, terrible typography, most accurate tool calling, and the only model to run an accessibility audit on its own page.
Takeaway: for web UI, a 27B model is already enough for professional output, and heavily quantized IQ1 still delivers surprisingly strong results. Full prompt and comparison table at basalt.jgulfo.com. Author discloses they maintain the engine and harness used.
More from coding & agent
- aivoicebus 0.16.0 adds Grok Bot to its AI agent fleet via webhook — baditaflorin · 2026-11-18
- deadweight: a Rust CLI that measures wasted tokens in your Claude Code setup — JeremyCMorgan · 2026-11-18
- Shadow MCP services need governance: 'no policy means you're asking for failures' — colinmcnamara · 2026-11-18
- LangChain's GTM Agent: deals with agent-built sales docs advance 47% vs 16% — LangChain · 2026-11-18
- LangChain case study: GTM agent artifacts moved 47% of deals to the next stage — LangChain · 2026-11-18
- Trader wires Muse agents to execute CME futures trades, plans broker API for instant orders — CyborgTraderOFTAGI · 2026-11-18