harness-hub/Qwen3.8-27B-opencode-adapted (revision b4171c0) merged into
Qwen/Qwen3.8-27B (revision 3756181) and converted to GGUF for llama.cpp:
ready-to-run files, no --lora needed. Every file keeps Qwen's multi-token-prediction (MTP) layer, so llama-server can
use it for faster decoding, and mmproj-Qwen3.8-27B-opencode-F16.gguf adds image input.
harness-hub trained this LoRA with reinforcement learning (AsyncGRPO in TRL) to be the model behind the OpenCode coding agent on software-engineering and terminal tasks. Their published results, under OpenCode with thinking off: SWE-bench Pro (hard 51) 35.3% → 51.0% and Terminal-Bench 2.1 46.4% → 52.4% against the base model, and with thinking off it matches the base model thinking at its highest effort while generating about a sixth of the tokens. See the original adapter card for the method, data and full results.
| File | Size | Same next token as BF16 | Mean KL divergence |
|---|---|---|---|
| Qwen3.8-27B-opencode-Q6_K.gguf (recommended) | 22.4 GB | 98.7% | 0.0197 |
| Qwen3.8-27B-opencode-Q8_0.gguf | 29.0 GB | 99.5% | 0.0002 |
| Qwen3.8-27B-opencode-Q4_K_M.gguf | 16.8 GB | 94.2% | 0.0689 |
| Qwen3.8-27B-opencode-Q3_K_M.gguf | 13.5 GB | 92.9% (vs Q8_0) | 0.0883 (vs Q8_0) |
| Qwen3.8-27B-opencode-IQ3_XXS.gguf | 11.4 GB | 91.0% (vs Q8_0) | 0.1356 (vs Q8_0) |
| Qwen3.8-27B-opencode-IQ2_M.gguf | 10.2 GB | 89.6% (vs Q8_0) | 0.1709 (vs Q8_0) |
| mmproj-Qwen3.8-27B-opencode-F16.gguf | 0.9 GB | vision projector |
Q6_K is the recommended balance of size and fidelity; Q8_0 is closest to the BF16 weights; Q4_K_M and the smaller Q3_K_M, IQ3_XXS, IQ2_M files are for machines with less memory. For Apple Silicon there are also MLX builds.
The adapter was trained and evaluated with thinking off, so keep it off:
hf download netprojects/Qwen3.8-27B-opencode-adapted-GGUF Qwen3.8-27B-opencode-Q6_K.gguf mmproj-Qwen3.8-27B-opencode-F16.gguf --local-dir .
llama-server -m Qwen3.8-27B-opencode-Q6_K.gguf -mm mmproj-Qwen3.8-27B-opencode-F16.gguf \
--jinja --chat-template-kwargs '{"enable_thinking": false}' \
--spec-type draft-mtp -c 65536 --port 8000 --alias qwen3.8-opencode
--spec-type draft-mtp turns on the built-in MTP layer; leave it out for plain decoding, and leave out -mm if you
do not need images. Then 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:8000/v1", "apiKey": "dummy" },
"models": { "qwen3.8-opencode": {} }
}
}
}
opencode run --model local/qwen3.8-opencode "Find out why the tests in tests/test_parser.py fail and fix the bug."
Measured with llama-server (Homebrew llama.cpp, commit b1b770d90) on a MacBook Pro M5 Max: greedy, thinking off, 5 prompts × 256 tokens from GSM8K, MATH-500, HumanEval, MBPP and MT-Bench, with other services running on the Mac, so treat the speeds as indicative.
| File | Plain | --spec-type draft-mtp |
Drafts accepted |
|---|---|---|---|
| Q4_K_M | 27 tok/s | 53 tok/s (×1.97) | 77% |
| Q6_K | 22 tok/s | 46 tok/s (×2.07) | 73% |
| Q8_0 | 18 tok/s | 43 tok/s (×2.36) | 74% |
These builds also run in Splash (1.7.0), which uses its own DFlash2 draft model for speculative decoding. Measured on the same 5 prompts × 256 tokens, thinking off, on 19 November 2026 with other models loaded on the Mac, so treat the speeds as indicative; text, tool calling and image input passed on every build:
| File | Decode speed | End to end, incl. prompt | Tokens per cycle |
|---|---|---|---|
| Q4_K_M | 144 tok/s | 111 tok/s | 6.6 |
| Q6_K | 129 tok/s | 98 tok/s | 6.9 |
| Q8_0 | 105 tok/s | 85 tok/s | 6.6 |
splash serve --model-dir Qwen3.8-27B-opencode-Q6_K.gguf --mmproj mmproj-Qwen3.8-27B-opencode-F16.gguf \
--model local/qwen3.8-opencode --reasoning-effort none
Q3_K_M, IQ3_XXS, IQ2_M were added on 19 November 2026. They were quantized from our Q8_0 file with an importance matrix (llama-imatrix on the Q8_0, English news text and MIT-licensed Python code, 512-token chunks). The MTP layer is kept at Q4_K in the IQ files, because the importance matrix does not cover it. Same next token and KL divergence for these files are measured against the Q8_0 file on the same 12 prompts. With llama-server (thinking off, with mmproj) each answered a short question correctly, produced a well-formed tool call and described the test image; three short Python tasks run against asserts: Q3_K_M 3/3 coding, IQ3_XXS 3/3 coding, IQ2_M 3/3 coding.
On a MacBook Pro M5 Max (128 GB), 18 November 2026.
W + 2.0 · B @ A rounded to bf16, checked element by element; every other tensor, including the MTP layer and
the vision encoder, is byte-identical to Qwen3.8-27B.bash and read tools produced a well-formed bash tool call, and a test
image was described correctly (red square, blue circle).Converted with llama.cpp's convert_hf_to_gguf.py at commit 8642541b2 and quantized with llama-quantize from Homebrew
llama.cpp (commit b1b770d90). The tokenizer and chat template are harness-hub's files, embedded unchanged.
The adapter is by harness-hub; the base model is Qwen3.8-27B by the Qwen Team, Alibaba Cloud. If you use the model, please cite their work (see the adapter card and the Qwen3.8-27B card).
Apache-2.0, as Qwen3.8-27B and the harness-hub adapter. LICENSE is the Apache-2.0 text from the Qwen3.8-27B repository
and NOTICE describes the conversion and credits harness-hub and the Qwen Team (Alibaba Cloud).
Published by NetProjects (Internet Projects Ltd). Not affiliated with Hugging Face, harness-hub or Alibaba.
More builds from NetProjects: huggingface.co/netprojects · netprojects.com
harness-hub/Qwen3.8-27B-opencode-adapted