Key Info

An experimental super-adapter for Alibaba's Qwen3.8-27B merges all of the week's harness LoRA adapters into a single larger LoRA by concatenating them, reporting accuracy gains and fewer tool calls.

Highlights

  • SWE-Bench Pro (hard 51): Hermes Agent 45% to 61%, OpenCode 47% to 67%, pi unchanged at 74%.
  • Terminal-Bench 2.1 gains +4 to +6 points under each harness.
  • In thinking mode, tool calls drop 19-48% depending on the harness.
  • The full recipe, experimental results and open questions are shared in a blogpost.