Qwen3.8-27B · opencode (LoRA, GGUF)
The GGUF version of harness-hub/Qwen3.8-27B-opencode-adapted, a LoRA adapter trained with reinforcement learning inside OpenCode, on Qwen/Qwen3.8-27B, for llama.cpp.
- SWE-bench Pro (hard 51): 35.3% → 51.0% (+15.7 points; 11 tasks newly solved, 3 lost).
- Terminal-Bench 2.1: 46.4% → 52.4% (+6.0 points); +13.6 points on the tasks not used to pick the checkpoint (p = 0.035).
- Less work per task: a quarter of the base model's turns on SWE-bench Pro and half on Terminal-Bench 2.1, on the tasks both solve.
Maximum-effort thinking accuracy, at a sixth of the tokens. With thinking off, the adapter matches the base model thinking at its highest reasoning effort (
xhigh), while generating 6–8× fewer tokens per task. That matters most locally, where every generated token is time: keep thinking off.
These scores come from the BF16 adapter served by vLLM, under the OpenCode 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-opencode-lora-F16.gguf is the adapter only (about 319 MB). It runs on top of the base model's GGUF.
Run
1. Download the base model and the adapter
hf download ggml-org/Qwen3.8-27B-GGUF Qwen3.8-27B-Q8_0.gguf --local-dir .
hf download harness-hub/Qwen3.8-27B-opencode-adapted-GGUF Qwen3.8-27B-opencode-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.
2. Serve
llama serve -m Qwen3.8-27B-Q8_0.gguf --lora Qwen3.8-27B-opencode-lora-F16.gguf \
--chat-template-kwargs '{"enable_thinking": false}' -c 65536 --port 8080 --alias qwen3.8-opencode
Keep thinking off, as in training and evaluation: with it on, the adapter is not significantly more accurate and generates about 6× more tokens.
3. Point OpenCode at it
Add the server to ~/.config/opencode/opencode.json:
{
"$schema": "https://opencode.ai/config.json",
"provider": {
"local": {
"npm": "@ai-sdk/openai-compatible",
"options": { "baseURL": "http://localhost:8080/v1", "apiKey": "dummy" },
"models": { "qwen3.8-opencode": {} }
}
}
}
opencode run --standalone --model local/qwen3.8-opencode "Find out why the tests in tests/test_parser.py fail and fix the bug."
Conversion
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
Limitations
- Not re-evaluated in GGUF. The scores on the original adapter card come from vLLM at BF16. This file has not been evaluated through llama.cpp, and a quantized base will move the results further.
- Thinking off. The adapter was trained without thinking; with thinking on it is not significantly more accurate, at about 6× the tokens.
- Vision and web search. Evaluated text only and without web search: OpenCode sessions that read an image or ask for a web search ended there. With a vision-capable base GGUF and web access, those tasks run normally.
- Downloads last month
- 1,229
16-bit
