VQV Signal

AI AT WORK WATCH TECHNICAL

Post on X reports one merged LoRA adapter lifts Qwen3.8-27B across three coding harnesses

According to a post on X by @yb2698, three nearly orthogonal RL-trained harness adapters were concatenated into a single rank-96 LoRA for Qwen3.8-27B. The author reports SWE-Bench Pro (hard 51) gains from 45% to 61% in Hermes Agent and from 47% to 67% in OpenCode, while pi stayed unchanged at 74%. The post also reports a gain of 4 to 6 points on Terminal-Bench 2.1 and 19–48% fewer tool calls with thinking on. The author says the weights and full recipe are public.

Source: Scobleizer News Pulse (X) · x.com Found via Robert Scoble’s Scobleizer News Pulse · a post on X, not independently verified Published 2026-11-20T14:45:00+00:00 Detected 2026-11-21T00:02:58+00:00
View original source

According to a post on X by @yb2698, three nearly orthogonal RL-trained harness adapters were concatenated into a single rank-96 LoRA for Qwen3.8-27B. The author reports SWE-Bench Pro (hard 51) gains from 45% to 61% in Hermes Agent and from 47% to 67% in OpenCode, while pi stayed unchanged at 74%. The post also reports a gain of 4 to 6 points on Terminal-Bench 2.1 and 19–48% fewer tool calls with thinking on. The author says the weights and full recipe are public.

AI-assisted summary based on the listed source.

@yb2698 concatenated three nearly orthogonal RL-trained harness adapters into one rank-96 LoRA. They report SWE-Bench Pro (hard 51) gains from 45% to 61% in Hermes Agent and from 47% to 67% in OpenCode (pi unchanged at 74%), +4 to +6 points on Terminal-Bench 2.1, and 19–48% fewer tool calls with thinking on. The...

Developers who run open models for coding agents may be able to tune one adapter that serves several harnesses, instead of maintaining separate ones, if these results hold. Because the weights and recipe are public, teams can check the reported numbers directly. Watch for independent reproductions of the SWE-Bench Pro and Terminal-Bench 2.1 results.

Teams using AI at work may want to compare this against current productivity and review workflows.

Signal Strength 92% Technical label WATCH Public Interest 32 Category AI AT WORK Reader Depth TECHNICAL

Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.

Public Interest components
Recognizable Entity Score 0 Practical Impact Score 46 Novelty Interest Score 70 Consequence Score 18 Curiosity Score 16 Shareability Score 49

VQV surfaced this signal because it is recent, relevant to Open Source LLMs, connected to Scobleizer News Pulse (X).