The GGUF version of harness-hub/Qwen3.8-27B-pi-adapted, a LoRA adapter trained with reinforcement learning inside pi, on Qwen/Qwen3.8-27B, for llama.cpp.
[!Tip] Maximum-effort thinking accuracy, at a quarter of the tokens. With thinking off, the adapter matches the base model thinking at its highest reasoning effort (
xhigh), while generating about 4.5× fewer tokens per task. That matters most locally, where every generated token is time: keep thinking off.
[!Note] These scores come from the BF16 adapter served by vLLM, under the pi harness (one attempt per task, a 100-step budget, thinking off, temperature 0.8). This GGUF file has not been re-evaluated through llama.cpp, and a quantized base will move the results further. They are not comparable with the scores on the Qwen3.8-27B card, which uses other harnesses and thinking at maximum effort.
The original adapter card describes the training and the evaluation.
Qwen3.8-27B-pi-lora-F16.gguf is the adapter only (about 319 MB). It runs on top of the base model's GGUF.
hf download ggml-org/Qwen3.8-27B-GGUF Qwen3.8-27B-Q8_0.gguf --local-dir .
hf download harness-hub/Qwen3.8-27B-pi-adapted-GGUF Qwen3.8-27B-pi-lora-F16.gguf --local-dir .
The adapter was trained on the BF16 weights. The closer the base is to them, the closer you get to the trained behavior.
Q8_0 is a good default; Q4_K_M is smaller, but drifts further from what the adapter learned against.
llama serve -m Qwen3.8-27B-Q8_0.gguf --lora Qwen3.8-27B-pi-lora-F16.gguf \
--chat-template-kwargs '{"enable_thinking": false}' -c 65536 --port 8080 --alias qwen3.8-pi
Keep thinking off, as in training and evaluation: with it on, the adapter is no more accurate and generates about 4× more tokens.
pi install git:github.com/huggingface/pi-llama
pi --model llama-cpp/qwen3.8-pi
Converted from the PEFT adapter with llama.cpp's convert_lora_to_gguf.py, at F16:
python llama.cpp/convert_lora_to_gguf.py <adapter dir> --base <Qwen/Qwen3.8-27B config dir> --outtype f16
harness-hub/Qwen3.8-27B-pi-adapted