NousResearch Releases Hermes Adapter for Qwen3.8-27B, Boosting SWE-Bench Pro to 43.1%
On November 19, NousResearch released its third lightweight Hermes adapter for Alibaba's Qwen3.8-27B (following the Pi and OpenCode versions), making the model a stronger and more efficient agent within the Hermes Agent ecosystem.
Confirmed
- SWE-Bench Pro (hard 51) pass rate improved from the base model's 29.4% to 43.1%, a 13.7-point gain, with roughly 3.5x savings in token consumption
- The adapter is a LoRA with r=32, about 320MB (F16 GGUF), trained on a frozen Qwen3.8-27B, with GGUF weights open-sourced under Apache-2.0
- Training recipe: every rollout was a real Hermes session in an HF sandbox, with training done via reinforcement learning
- Efficiency comparison: under the same 10k token per-task budget, the adapter solved 51% of TB-2.1 tasks and 39% of SWE-Bench Pro tasks, while the base model at the xhigh setting solved only 28% (TB-2.1)
- The author added a real SWE-Bench Pro task demo: the base model broke a test file and exhausted its 100-step budget without solving it, while the adapter version solved it in just 35 tool calls
- According to @ariG23498, the adapter maintains the 43.1% score even in no-thinking mode
Why it matters
- A tiny LoRA (320MB) trained with RL on a specific harness delivers major agent capability gains and lower token costs, showcasing the cost-effectiveness of the lightweight adapter approach
- RL training driven by real sessions plus open weights (Apache-2.0) lets the community directly reproduce and deploy the work
2026-11-19 ~ 2026-11-19 · 11 related posts
Primary sources
- 320MB LoRA RL-trained in Hermes Agent gives Qwen3.8-27B thinking-level accuracy with 3.5–4.4× fewer tokens — HauntingMoment ·
- NousResearch's Hermes adapter lifts Qwen3.8-27B agent scores: SWE-Bench Pro 29.4%→43.1% at 4× fewer tokens — yb2698 ·
- RL-Trained Hermes LoRA on Qwen3.8-27B Adds +13.7 Pts on SWE-Bench Pro While Using Fewer Tokens — yb2698 ·
- [source] NousResearch's Hermes adapter lifts Qwen3.8-27B agent scores: SWE-Bench Pro 29.4%→43.1% at 4× fewer tokens — yb2698 · 2026-11-19
- Hermes adapter solves a SWE-Bench Pro task in 35 tool calls where the base Qwen3.8 model failed — yb2698 · 2026-11-19
- Same 10k-Token Budget: Hermes Adapter Solves 51% of TB-2.1 vs Base Model's 28% with xhigh Thinking — yb2698 · 2026-11-19
- [source] RL-Trained Hermes LoRA on Qwen3.8-27B Adds +13.7 Pts on SWE-Bench Pro While Using Fewer Tokens — yb2698 · 2026-11-19
- [source] 320MB LoRA RL-trained in Hermes Agent gives Qwen3.8-27B thinking-level accuracy with 3.5–4.4× fewer tokens — HauntingMoment · 2026-11-19
- NousResearch ships Hermes adapter, lifting Qwen3 SWE-Bench Pro from 29.4% to 43.1% — ariG23498 · 2026-11-19
- Same SWE-Bench task: base model burns 100-step budget, adapter passes in 35 calls — yb2698 · 2026-11-19
- Within 10k tokens, Hermes adapter solves 51% of TB-2.1 vs base's 28% — yb2698 · 2026-11-19
- NousResearch Hermes adapter lifts Qwen3.8-27B non-thinking perf 14 points, cuts tokens 3.5-4.4x — yb2698 · 2026-11-19