harness-hub's Qwen3.8-27B super-adapter on a Mac: one model for pi, Hermes Agent and OpenCode
NetProjects
harness-hub's super-adapter is one LoRA adapter for Alibaba Cloud's Qwen3.8-27B that sits behind three coding agents: pi, Hermes Agent and OpenCode. harness-hub trained three specialists with reinforcement learning, one inside each agent, on software-engineering and terminal tasks, then merged them at a weight of one half each (pi and Hermes Agent at step 160, OpenCode at step 150). The merge itself involves no further training.
harness-hub recommends using it with thinking on. Run under each agent on SWE-bench Pro (hard 51) and Terminal-Bench 2.1, it reports mean pass@1 rising from 55.5% to 63.5% against the base model thinking at the same effort, ahead on five of the six agent and benchmark pairs and level on the sixth, with 19 to 48% fewer tool calls per task. On SWE-bench Pro it reports 47.1% to 66.7% under OpenCode and 45.1% to 60.8% under Hermes Agent. With thinking off, it reports the adapter ahead on all six pairs, mean 39.5% to 49.0%, with 53 to 73% fewer tool calls. These are harness-hub's figures, from its own setup; we have not reproduced them.
We merged the adapter into Qwen3.8-27B, kept the vision encoder and Qwen's own multi-token-prediction (MTP) layer, and published five ready-to-run builds for Macs with Apple Silicon (MLX) and a GGUF repository for llama.cpp. Everything is under the Apache-2.0 licence. If you use a single agent with thinking off, harness-hub recommends that agent's specialist instead; we publish those too: https://netprojects.com/insights/harness-hub-qwen3-8-27b-pi-adapted-coding-agent-mlx-gguf
Always up to date. We keep these builds current on Hugging Face. When harness-hub releases a new version of the adapter, we rebuild, re-test and publish it there, so the links on this page always point to the latest version. The numbers in this article were measured on the version named below; the model cards on Hugging Face carry the current figures.
- Hugging Face: https://huggingface.co/netprojects
- NetProjects MLX builds collection: https://huggingface.co/collections/netprojects/netprojects-mlx-builds
What we released
Merged from harness-hub's adapter at revision adb00e9 and Qwen3.8-27B at revision 3756181. The MLX builds were converted with mlx-vlm and use affine quantisation with group 64; in the 3-bit build the token embedding, output layer and the more sensitive layers stay at 4-bit, and the 2.5-bit build is text only. Peak memory is for an 8K-token prompt with mlx-lm (text). The bf16 row is our local reference conversion, which we have not published.
| Hugging Face | Size | Same next token as bf16 | GSM8K, 50 questions | Peak memory | Suggested Mac memory |
|---|---|---|---|---|---|
| netprojects/Qwen3.8-27B-super-adapter-MLX-6bit (recommended) | 23.1 GB | 99.5% | 45/50 | 23.4 GB | 36 GB+ |
| netprojects/Qwen3.8-27B-super-adapter-MLX-8bit | 30.0 GB | 99.6% | 45/50 | 29.7 GB | 48 GB+ |
| netprojects/Qwen3.8-27B-super-adapter-MLX-4bit | 16.3 GB | 96.2% | 45/50 | 17.1 GB | 32 GB+ |
| netprojects/Qwen3.8-27B-super-adapter-MLX-3bit | 13.8 GB | 92.2% | 46/50 | 14.6 GB | 24 GB+ |
| netprojects/Qwen3.8-27B-super-adapter-MLX-2.5bit (text only) | 11.8 GB | 85.2% | 41/50 | 13.8 GB | 24 GB+ |
| bf16 reference (not published) | 55.6 GB | reference | 45/50 | 53.4 GB | 96 GB+ |
The GGUF repository, netprojects/Qwen3.8-27B-super-adapter-GGUF, has Q6_K (22.4 GB, recommended), Q8_0 (29.0 GB) and Q4_K_M (16.8 GB), each with the MTP layer, plus a 0.9 GB vision projector for image input. Compared with the BF16 GGUF, they pick the same next token on 98.9%, 99.7% and 94.6% of positions.
How well it matches the original
Measured on 21 November 2026, MacBook Pro M5 Max (128 GB), with mlx 0.32.4, mlx-lm 0.32.2 and mlx-vlm 0.7.11.
- Merge: all 256 adapted modules equal the base weight plus 2.0 times the LoRA product, rounded to bf16, checked element by element against an independent computation. Every other tensor, including the vision encoder and the MTP layer, is the same as in Qwen3.8-27B, byte for byte. On 24 sampled modules, the merged change equals half the sum of the three specialists' changes.
- Quantisation: each quantised MLX build is exactly the affine (group 64) quantisation of the bf16 build.
- Agreement with bf16, on 12 short teacher-forced prompts (946 positions for the 8-bit, 6-bit and 4-bit builds): 8-bit 99.6%, 6-bit 99.5%, 4-bit 96.2%, 3-bit 92.2%, 2.5-bit 85.2%.
- Tool calling: with the included chat template and thinking off, every MLX build and each GGUF file produced a well-formed bash tool call on an agent-style prompt.
- Coding: three short Python tasks run against asserts, 3 of 3 on the 8-bit, 6-bit, 4-bit and 3-bit MLX builds and on each GGUF file, 2 of 3 on the 2.5-bit MLX build.
- Thinking on: every MLX build answered two short reasoning questions correctly, each after one closed thinking block, with no looping.
- Vision: the 8-bit, 6-bit, 4-bit and 3-bit MLX builds and the GGUF files with the vision projector described a test image correctly.
- MTP: on the 4-bit MLX build, greedy decoding went from 33.1 to 65.1 tokens per second with the MTP layer; on the Q4_K_M GGUF file with llama-server, from 28.4 to 51.5. Both from five short prompts of 256 tokens with thinking off, on a Mac with other work running, so treat them as indicative.
- Source files: every file of both pinned upstream revisions matches the Hub's SHA-256 hashes.
Try it
harness-hub recommends thinking on, which is the chat template's default. It evaluated thinking-on runs with up to 32,768 output tokens per model call, so allow long replies. On a Mac with Apple Silicon, serve it with mlx-lm's OpenAI-compatible server and point pi, Hermes Agent or OpenCode at it; Hermes Agent needs at least 64,000 tokens of context.
pip install -U "mlx-lm>=0.32.2"
mlx_lm.server --model netprojects/Qwen3.8-27B-super-adapter-MLX-6bit --port 8000The model card has the configuration for all three agents, image input with mlx-vlm and the MTP command: https://huggingface.co/netprojects/Qwen3.8-27B-super-adapter-MLX-6bit. The GGUF card has the llama.cpp commands: https://huggingface.co/netprojects/Qwen3.8-27B-super-adapter-GGUF
Honest limits
- We did not run SWE-bench Pro, Terminal-Bench or any other agent test suite on these builds, so the agent scores above are harness-hub's, from its own setup, text only.
- harness-hub notes that each benchmark has few tasks, so most single results are within noise, and that the merge weight was chosen on Terminal-Bench 2.1, so its Terminal-Bench results are slightly optimistic. The specialists were trained with thinking off; thinking on was evaluated, not trained.
- Our checks are short: 12 prompts for agreement, one tool-calling prompt, three coding tasks, two reasoning questions and 50 GSM8K questions, all on one machine.
- Peak memory is for an 8K-token prompt; Hermes Agent's 64,000-token context needs more.
- The 3-bit and 2.5-bit builds differ noticeably from the bf16 weights, and the 2.5-bit build has no image input.
Licence
Qwen3.8-27B, harness-hub's adapter and our builds are all under the Apache-2.0 licence. Each of our repositories carries the licence text and a notice that describes the conversion and credits harness-hub and the Qwen team (Alibaba Cloud).
Links
- MLX 6-bit build (recommended): https://huggingface.co/netprojects/Qwen3.8-27B-super-adapter-MLX-6bit
- MLX 8-bit build: https://huggingface.co/netprojects/Qwen3.8-27B-super-adapter-MLX-8bit
- MLX 4-bit build: https://huggingface.co/netprojects/Qwen3.8-27B-super-adapter-MLX-4bit
- MLX 3-bit build: https://huggingface.co/netprojects/Qwen3.8-27B-super-adapter-MLX-3bit
- MLX 2.5-bit build (text only): https://huggingface.co/netprojects/Qwen3.8-27B-super-adapter-MLX-2.5bit
- GGUF builds: https://huggingface.co/netprojects/Qwen3.8-27B-super-adapter-GGUF
- harness-hub's original adapter: https://huggingface.co/harness-hub/Qwen3.8-27B-super-adapter
- pi: https://github.com/earendil-works/pi
- Hermes Agent: https://github.com/NousResearch/hermes-agent
- OpenCode: https://opencode.ai
- Qwen3.8-27B: https://huggingface.co/Qwen/Qwen3.8-27B
- All our AI ports: https://netprojects.com/ai
Credits
The adapter is by harness-hub; Qwen3.8-27B is by Alibaba Cloud's Qwen team, under the Apache-2.0 licence. The MLX and GGUF builds are by NetProjects (Internet Projects Ltd). Not affiliated with harness-hub or Alibaba Cloud.
This article was drafted with AI assistance, then edited and approved by people on our team. You can read more about how we use AI at https://netprojects.com/standards/ai.