harness-hub/Qwen3.8-27B-hermes-adapted (revision 09fb741) merged into
Qwen/Qwen3.8-27B (revision 3756181) and converted to MLX with mlx-vlm, vision
encoder and Qwen's MTP layer included. In this build most linear layers are quantized to 3-bit (affine, group 64); the token embedding, the output layer and the more sensitive v_proj / down_proj layers (the first and last eight layers and every third layer in between) are 4-bit, about 3.91 bits per weight overall. Size: 13.8 GB.
harness-hub trained this LoRA with reinforcement learning (AsyncGRPO in TRL, LoRA rank 32, alpha 64, checkpoint at step 160 of 1000 planned) to be the model behind Hermes Agent on software-engineering and terminal tasks. Their published results, under Hermes Agent with thinking off: SWE-bench Pro (hard 51) 29.4% → 43.1% and Terminal-Bench 2.1 44.7% → 58.8% against the base model, using under 40% of the base model's turns on the tasks both solve and about 2× fewer generated tokens. With thinking off it is level with the base model thinking at its highest effort (xhigh), at 3.5–4.4× fewer tokens. See the original adapter card for the method, data and full results.
| Build | Size | Same next token as bf16 | KL vs bf16 | GSM8K (50 questions) | Peak memory, 8K-token prompt | Suggested Mac memory |
|---|---|---|---|---|---|---|
| netprojects/Qwen3.8-27B-hermes-adapted-MLX-6bit (recommended) | 23.1 GB | 98.5% | 0.0015 | 46/50 | 23.4 GB | 36 GB+ |
| netprojects/Qwen3.8-27B-hermes-adapted-MLX-8bit | 30.0 GB | 99.2% | 0.0004 | 45/50 | 29.7 GB | 48 GB+ |
| netprojects/Qwen3.8-27B-hermes-adapted-MLX-4bit | 16.3 GB | 96.1% | 0.0192 | 43/50 | 17.1 GB | 32 GB+ |
| netprojects/Qwen3.8-27B-hermes-adapted-MLX-3bit | 13.8 GB | 91.7% | 0.0824 | 45/50 | 14.6 GB | 24 GB+ |
| netprojects/Qwen3.8-27B-hermes-adapted-MLX-2.5bit | 11.8 GB | 85.4% | 0.2461 | 44/50 | 13.8 GB | 24 GB+ |
| netprojects/Qwen3.8-27B-hermes-adapted-MLX-bf16 | 55.6 GB | reference | reference | 45/50 | 53.4 GB | 96 GB+ |
| netprojects/Qwen3.8-27B-hermes-adapted-MLX-adapter | 638 MB | LoRA only, for mlx-lm | on any Qwen3.8-27B MLX build |
The 6-bit build is the recommended balance of size and fidelity; 8-bit is closest to bf16; 4-bit, 3-bit and the text-only 2.5-bit are for Macs with less memory (3-bit and 2.5-bit fit 24 GB Macs). The columns are explained under "What was checked".
The adapter was trained and evaluated with thinking off, so keep it off. Serve the model with mlx-lm's OpenAI-compatible server:
pip install -U "mlx-lm>=0.32.2"
mlx_lm.server --model netprojects/Qwen3.8-27B-hermes-adapted-MLX-3bit --port 8000 \
--chat-template-args '{"enable_thinking": false}'
Install Hermes Agent:
curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash
Add the server to ~/.hermes/config.yaml:
model:
default: netprojects/Qwen3.8-27B-hermes-adapted-MLX-3bit
provider: custom
base_url: http://localhost:8000/v1
api_key: dummy
Hermes Agent needs at least 64,000 tokens of context. Then:
hermes chat -q "Find out why the tests in tests/test_parser.py fail and fix the bug."
Text only, with mlx-lm:
from mlx_lm import load, generate
model, tokenizer = load("netprojects/Qwen3.8-27B-hermes-adapted-MLX-3bit")
messages = [{"role": "user", "content": "Write a Python function that parses '1.5M' into bytes."}]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, enable_thinking=False, tokenize=False)
print(generate(model, tokenizer, prompt=prompt, max_tokens=512))
With images, with mlx-vlm:
pip install -U "mlx-vlm>=0.7.11"
mlx_vlm.generate --model netprojects/Qwen3.8-27B-hermes-adapted-MLX-3bit --image screenshot.png \
--prompt "What error does this terminal show?" --max-tokens 300 # thinking stays off unless --enable-thinking
The tokenizer and chat template are harness-hub's files, unchanged (same vocabulary and tool-call format as Qwen3.8-27B).
Tool calls use Qwen's XML format (<tool_call><function=...><parameter=...>), which mlx-lm's server returns as
OpenAI tool_calls.
This build keeps Qwen3.8-27B's own multi-token-prediction (MTP) layer in model-mtp.safetensors, quantized to 8-bit and declared in config.json as mtplx_mtp_quantization. The adapter does not
change that layer, so it is Qwen's original MTP head. mlx-vlm uses it as a built-in draft model when you pass the same
repository as --draft-model:
mlx_vlm.generate --model netprojects/Qwen3.8-27B-hermes-adapted-MLX-3bit --draft-model netprojects/Qwen3.8-27B-hermes-adapted-MLX-3bit \
--draft-kind mtp --temperature 0 --max-tokens 512 --prompt "Write a Python function that checks whether a number is prime."
Every drafted token is checked by the full model. With greedy decoding the text matches plain decoding until a near-tie, where the batched check can round the other way, so long answers can diverge slightly.
On a MacBook Pro M5 Max (128 GB), 19 November 2026, with mlx 0.32.4, mlx-lm 0.32.2 and mlx-vlm 0.7.11.
W + 2.0 · B @ A rounded to bf16, checked element by element against an independent computation; every other tensor is byte-identical to Qwen3.8-27B. The adapter does not touch the vision encoder or the Gated DeltaNet layers.bash and read tools; 8-bit, 6-bit, 4-bit, 3-bit, 2.5-bit each produced a well-formed bash tool call in Qwen's XML format. Three short Python tasks run against asserts: 8-bit 3/3, 6-bit 3/3, 4-bit 3/3, 3-bit 3/3, 2.5-bit 2/3.sudo sysctl iogpu.wired_limit_mb=...), leaving several GB for macOS, and use short contexts.mtp.* weights (norms in mlx-vlm's stored form), quantized like the rest of the build; the 4-bit, 3-bit and 2.5-bit builds were loaded in mlx-lm and mlx-vlm with the extra file present and decoded with its bundled MTP head.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 MLX builds from NetProjects: huggingface.co/netprojects · netprojects.com
model-00002-of-00003.safetensors 5.4 GBmodel-00001-of-00003.safetensors 5.4 GBmodel-00003-of-00003.safetensors 2.6 GBmodel-mtp.safetensors 451 MBharness-hub/Qwen3.8-27B-hermes-adapted