Same 10k-Token Budget: Hermes Adapter Solves 51% of TB-2.1 vs Base Model's 28% with xhigh Thinking
yb2698 · x · 2026-11-19
Follow-up data on the Qwen3.8-27B Hermes adapter, comparing thinking vs no-thinking: within a 10k-token budget per task, the adapter solves 51% of TB-2.1 and 39% of SWE-Bench Pro, while the base model at xhigh thinking manages only 28% and 0%. Median tool calls on commonly solved tasks drop from 23 to 7 (TB-2.1) and 63 to 28 (SWE-Bench Pro). The author also demos a real SWE-Bench Pro task where the base model breaks the test file and exhausts its 100-step budget, while the adapter fixes the bug and passes in 35 tool calls.
Related event: Hermes Adapter Beats Base Model in Real-World Benchmarks(2 posts)→
More from coding & agent
- Qwen3.8-27B + Hermes adapter lifts SWE-bench Pro to 43% with 4x fewer tokens — yb2698 · 2026-11-19
- Every on Claude Fable 5.2: Human Thinking at the Start Matters More as AI Builds More — every · 2026-11-19
- Semantic caching LLM answers by embedding similarity is unsafe — 138-pair test shows why — Narrow_Side_7778 · 2026-11-19
- Can ComfyUI refine existing mocap animations? A motion-to-motion workflow question — _Tito5_ · 2026-11-19
- Mycelium: an open-source framework that turns analytical project knowledge into a living dataset — arjunrajlab · 2026-11-19
- harness-hub open-sources RL-trained LoRA adapters tuned per agent harness, recipe coming Friday — yb2698 · 2026-11-19