Skip to content
NetProjects
Menu

Insights

Nous Research's Arete-27B on a Mac: six MLX builds with an MTP head

NetProjects

Arete-27B is Nous Research's full supervised fine-tune of Alibaba Cloud's Qwen3.8-27B. Nous Research says it trained the model to give factual, direct answers on politically sensitive topics in seven areas, such as freedom of expression, democracy and data privacy, where it saw refusals, deflections and state-framed answers from the base model, while keeping its general and agentic skills. The training data is text only and the vision encoder was frozen.

On Nous Research's own evaluation of 803 prompts, the mean fact score rose from 2.32 to 2.98 out of 5 and the share of answers containing propaganda fell from 22% to 0.9%, with thinking off. Its general-capability test, Hermes Bench, moved from 65.77% to 65.60%, which Nous Research reads as noise, and Terminal-Bench 2.1 from 67% to 70%. These are Nous Research's figures for its original weights; we have not reproduced them.

We converted Arete-27B to MLX for Macs with Apple Silicon in six sizes, from bf16 down to a text-only 2.5-bit build, and kept the vision encoder. Arete-27B ships without a multi-token-prediction (MTP) head, so we added Qwen3.8-27B's own, unchanged, for faster decoding. Everything is under the Apache-2.0 licence.

Always up to date. We keep these builds current on Hugging Face. When Nous Research releases a new version of the model, 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 in each section; the model cards on Hugging Face carry the current figures.

What we released

Converted from NousResearch/Arete-27B at revision 108f0afb, using affine quantisation with group 64; the vision encoder stays in bf16. In the 3-bit build the token embedding, output layer and the more sensitive layers stay at 4-bit; the 2.5-bit build is text only. The MTP head is Qwen3.8-27B's, from revision 37561816.

Hugging FaceSizeSame next token as bf16GSM8K, 50 questions
netprojects/Arete-27B-MTP-MLX-6bit (recommended)23.1 GB98.7%46/50
netprojects/Arete-27B-MTP-MLX-8bit29.9 GB99.1%not run
netprojects/Arete-27B-MTP-MLX-4bit16.3 GB95.5%47/50
netprojects/Arete-27B-MTP-MLX-3bit13.8 GB90.1%49/50
netprojects/Arete-27B-MTP-MLX-2.5bit (text only)11.8 GB84.6%46/50
netprojects/Arete-27B-MTP-MLX-bf1655.6 GBreferencenot run

How well it matches the original

Measured on 19 November 2026, MacBook Pro M5 Max (128 GB), with mlx 0.32.4, mlx-lm 0.32.2 and mlx-vlm 0.7.11.

  • Source files: every file of Nous Research's pinned revision matches the Hub's SHA-256 hashes.
  • Quantisation: each quantised build is exactly the affine (group 64) quantisation of the bf16 build.
  • Agreement with bf16, over 953 teacher-forced positions: 8-bit 99.1%, 6-bit 98.7%, 4-bit 95.5%, 3-bit 90.1%, 2.5-bit 84.6%.
  • Tool calling and coding, thinking off: every quantised build produced a correct tool call on an agent-style prompt; three short Python tasks run against asserts were 3 of 3 on the 8-bit, 6-bit, 4-bit and 3-bit builds and 2 of 3 on the 2.5-bit build.
  • Thinking: on two short maths prompts every quantised build closed its thinking once and gave the right answer.
  • Vision: the 8-bit, 6-bit, 4-bit and 3-bit builds named the shapes and colours in a test image correctly.
  • Peak memory for an 8K-token prompt: 14.6 GB for the 3-bit build and 13.8 GB for the 2.5-bit build, so both fit 24 GB Macs.

Speed

Greedy decoding with mlx-vlm on the same Mac, same prompt set. The MTP head was trained against Qwen3.8-27B, not against Arete-27B, so it is a base-model head on a fine-tune; every drafted token is still checked by the full model.

MLX buildPlainWith MTPDraft acceptance
4-bit20.5 tok/s39.8 tok/s (×1.95)75%
8-bit13.1 tok/s28.6 tok/s (×2.18)77%

Try it

Nous Research evaluated Arete-27B with thinking off, so the example turns it off. On a Mac with Apple Silicon:

pip install -U "mlx-lm>=0.32.2"
mlx_lm.server --model netprojects/Arete-27B-MTP-MLX-6bit --port 8000 \
  --chat-template-args '{"enable_thinking": false}'

The model card has image input with mlx-vlm and the MTP command: https://huggingface.co/netprojects/Arete-27B-MTP-MLX-6bit

Honest limits

  • We did not run Nous Research's propaganda evaluation, Hermes Bench or Terminal-Bench on these builds, so the scores above are Nous Research's, from its own setup.
  • Our checks are short: one set of prompts for agreement, one tool-calling prompt, three coding tasks, two thinking prompts and 50 GSM8K questions; GSM8K was not run on the 8-bit and bf16 builds.
  • Only the 4-bit and 8-bit builds were timed, on one machine; speed-ups depend on the prompt and may be smaller on other text.
  • We measured peak memory only for the 3-bit and 2.5-bit builds.
  • 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

Arete-27B, Qwen3.8-27B 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.

Links

Credits

Arete-27B is by Nous Research; Qwen3.8-27B and its MTP head are by Alibaba Cloud's Qwen team, both under the Apache-2.0 licence. The MLX builds are by NetProjects (Internet Projects Ltd). Not affiliated with Nous Research 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.