Skip to content
NetProjects
Menu

Insights

RadixArk's DaisyRose-27B on a Mac: six MLX builds with image input and MTP

NetProjects

DaisyRose-27B is RadixArk's first model: a reasoning model built on Alibaba Cloud's Qwen3.8-27B with reinforcement learning only, with no supervised fine-tuning or distillation. RadixArk says it was trained on about 1,200 long-horizon agent tasks, for 18 days on 144 H200 GPUs, and that it keeps the base model's architecture, tokeniser, chat template, image understanding and 256K-token context. It thinks before it answers by default.

RadixArk reports gains over Qwen3.8-27B on most of the coding, reasoning and instruction-following tests it ran, for example Terminal Bench 4.1 from 5.6 to 12.0, Terminal Bench 2.1 from 73.0 to 75.5, DeepSWE v1.1 from 42.2 to 46.2 and IFBench from 79.5 to 82.4. These are RadixArk's figures for its original BF16 and FP8 weights; we have not reproduced them.

We converted DaisyRose-27B to MLX for Macs with Apple Silicon in six sizes, from bf16 down to a text-only 2.5-bit build, keeping the vision encoder and DaisyRose's own multi-token-prediction (MTP) layer. Everything is under the Apache-2.0 licence.

Always up to date. We keep these builds current on Hugging Face. When RadixArk 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 RadixArk/DaisyRose-27B at revision 293a7a2 with mlx-vlm, using affine quantisation with group 64. 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. Peak memory is for an 8K-token prompt with mlx-lm (text).

Hugging FaceSizeSame next token as bf16GSM8K, 50 questionsPeak memory
netprojects/DaisyRose-27B-MLX-6bit (recommended)23.1 GB98.7%44/5023.4 GB
netprojects/DaisyRose-27B-MLX-8bit29.9 GB99.5%45/5029.7 GB
netprojects/DaisyRose-27B-MLX-4bit16.3 GB95.8%45/5017.1 GB
netprojects/DaisyRose-27B-MLX-3bit13.8 GB92.1%43/5014.6 GB
netprojects/DaisyRose-27B-MLX-2.5bit (text only)11.8 GB87.6%42/5013.8 GB
netprojects/DaisyRose-27B-MLX-bf1655.6 GBreferencenot run53.4 GB

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 RadixArk'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, with the bit widths described on its model card.
  • Agreement with bf16, over 953 teacher-forced positions from 12 short prompts (9 with thinking off, 3 with thinking on): 8-bit 99.5%, 6-bit 98.7%, 4-bit 95.8%, 3-bit 92.1%, 2.5-bit 87.6%.
  • 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.
  • GSM8K, 50 fixed questions, thinking off, greedy: between 42 and 45 out of 50 across the quantised builds.
  • Vision: every build with vision named the shapes and colours in a test image correctly.
  • MTP: the bundled MTP layer equals DaisyRose-27B's own, and every quantised build decoded with it in mlx-lm and mlx-vlm.

Try it

On a Mac with Apple Silicon, serve it with mlx-lm's OpenAI-compatible server. RadixArk recommends temperature 1.0, top_p 0.95 and top_k 20 for thinking mode; the chat template also accepts enable_thinking set to false to answer directly.

pip install -U "mlx-lm>=0.32.2"
mlx_lm.server --model netprojects/DaisyRose-27B-MLX-6bit --port 8000 --temp 1.0 --top-p 0.95 --top-k 20

The model card has image input with mlx-vlm and the MTP command for faster decoding: https://huggingface.co/netprojects/DaisyRose-27B-MLX-6bit. The 6-bit build needs a Mac with 36 GB or more by default; the 3-bit and 2.5-bit builds fit 24 GB Macs.

Honest limits

  • We did not run Terminal Bench, DeepSWE, SWE Bench Pro or any other agent test suite on these builds, so the scores above are RadixArk's, from its own setup.
  • Our checks are short: 12 prompts for agreement, one tool-calling prompt, three coding tasks, two thinking prompts and 50 GSM8K questions.
  • All figures come from one machine. We did not time these builds, so we give no speed figures, with or without MTP.
  • Peak memory is for an 8K-token prompt; longer contexts need 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

DaisyRose-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 and credits RadixArk and the Qwen team (Alibaba Cloud).

Links

Credits

DaisyRose-27B is by RadixArk; Qwen3.8-27B is 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 RadixArk or Alibaba.

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.