AnalysisAI ModelsNovember 20, 2026

Qwen3.8-27B "super adapter" merges three agent harness LoRAs into one

Read original source →x.com

A single rank-96 LoRA stacking the weight updates from Hermes Agent, OpenCode and pi lifted SWE-Bench Pro from 55% to 63% mean, with 53–73% fewer tool calls. The adapters turned out to be nearly orthogonal, so concatenating them beat the base thinking model across harnesses.

2 sources

More stories today

Open the live feed