@Brjen’s team spent a week comparing local coding-agent stacks on one RTX 3090 (24GB). Winner: Qwen3.8-27B Unsloth UD-Q4_K_XL with llama.cpp (65k context, q8 KV, native MTP depth 2, fully on GPU). Real multi-step coding runs hit ~59 tok/s (~46 tok/s at 45k context) and passed nested tool JSON, fixing a failing test, and a clean unified diff. They rejected Q3 for bad tools, heavier quants for context/MTP tradeoffs, and Flash-Next setups that needed CPU offload.

Key Takeaways

  • ✓Winner: Qwen3.8-27B UD-Q4_K_XL + llama.cpp, 65k context and MTP, fully on a 3090.
  • ✓~59 tok/s on multi-step coding; passed tool JSON, test-fix, and correct diff gates.
  • ✓Q3, heavier quants, and offloaded Flash-Next were less reliable on 24GB.
ADSponsored