← sources

[Release] Qwen3.8-27B LoRA RL-trained inside Hermes Agent: thinking-level accuracy with thinking off, 3.5–4.4× fewer tokens, GGUF included

source: reddit (reddit:LocalLLaMA)published 2026-11-19 14:00ingested 2026-11-20 02:39

retained text

We RL-trained a small LoRA adapter (r=32, \~320 MB as an F16 GGUF) on a frozen Qwen3.8-27B, inside [Hermes Agent](https://github.com/NousResearch/hermes-agent). It's our third harness adapter, after pi and OpenCode.

**Results** (under Hermes Agent, one attempt per task, base and adapter on the same tasks and settings):

|Benchmark|Base|Base, thinking (xhigh)|Hermes adapter|Hermes adapter, thinking|
|:-|:-|:-|:-|:-|
|SWE-bench Pro (hard 51)|29.4%|45.1%|**43.1%**|51.0%|
|Terminal-Bench 2.1 (85 tasks)|44.7%|58.8%|**58.8%**|57.6%|

With thinking off, the adapter solves 19 Terminal-Bench tasks the base misses and loses 7 (p = 0.03). On SWE-bench Pro it's 10 newly solved and 3 lost.

**Run it locally with llama.cpp:**

    hf download ggml-org/Qwen3.8-27B-GGUF Qwen3.8-27B-Q8_0.gguf --local-dir .
    hf download harness-hub/Qwen3.8-27B-hermes-adapted-GGUF Qwen3.8-27B-hermes-lora-F16.gguf --local-dir .
    llama serve -m Qwen3.8-27B-Q8_0.gguf --lora Qwen3.8-27B-hermes-lora-F16.gguf \
            --chat-template-kwargs '{"enable_thinking": false}' -c 65536 --port 8080 --alias qwen3.8-hermes

Then point Hermes Agent at http://localhost:8080/v1. The config snippet is in the model card, along with vLLM, transformers serve, and merging into the base with PEFT.

Every evaluation trajectory, solved or not, is in a bucket linked from the card. The full recipe is in the blog post on Friday.

* Model: [https://huggingface.co/harness-hub/Qwen3.8-27B-hermes-adapted](https://huggingface.co/harness-hub/Qwen3.8-27B-hermes-adapted)
* GGUF: [https://huggingface.co/harness-hub/Qwen3.8-27B-hermes-adapted-GGUF](https://huggingface.co/harness-hub/Qwen3.8-27B-hermes-adapted-GGUF)
* OpenCode adapter: [https://huggingface.co/harness-hub/Qwen3.8-27B-opencode-adapted](https://huggingface.co/harness-hub/Qwen3.8-27B-opencode-adapted)
* pi adapter: [https://huggingface.co/harness-hub/Qwen3.8-27B-pi-adapted](https://huggingface.co/harness-hub/Qwen3.8-27B-pi-adapted)

Happy to answer questions!

document fetched 2026-11-20 00:42 · canonical https://reddit.com/r/LocalLLaMA/comments/1y0a1zy/release_qwen3827b_lora_rltrained_inside_hermes