Fine-tuned Qwen3.5-4B router model trained on 17.6k rows for about $5.50 on one A100
MaziyarPanahi · x · 2026-11-07
MaziyarPanahi released Qwen3.5-4B-Router-1M, a decision fine-tune of Alibaba's Qwen3.5-4B done with Unsloth on just 17,600 training rows. Training ran on Hugging Face Jobs with a single A100 GPU at a cost of about $5.50; the model is open on Hugging Face — a showcase of ultra-cheap small-model fine-tuning for routing decisions.
More from coding & agent
- Solo founder builds an AI sales agent that found 400+ leads in one run, 22 hot — RainbowFatDragon · 2026-11-07
- dotey ships DeepSeek Harness plugin update, has to build his own auto-update mechanism — dotey · 2026-11-07
- Managed Deep Agents can now be steered mid-turn while agents work — hwchase17 · 2026-11-07
- Every app should have an MCP server, argues Dan Vega — therealdanvega · 2026-11-07
- Google: stale TODO comments derail 1 in 7 coding agent sessions — rseroter · 2026-11-07
- Dev turns Runway AI video into a playable Unreal Engine game via agent — davidfromkansas · 2026-11-07