Qwen Concept 27B for Apple Silicon (MLX, 6-bit)
An MLX conversion of ConceptCountry/qwen-concept-27b
(revision 4c96373, bf16 weights). Qwen Concept 27B is Qwen3.8-27B with nine MLP down-projections edited by
Concept Country so that the model reaches its answer with less thinking
(paper). This build quantises the language model to 6-bit (affine, group 64) and keeps the vision
tower in bf16, so it runs in mlx-lm (text) and mlx-vlm (text and images). Size: 22.8 GB.
Published by NetProjects (Internet Projects Ltd). Not affiliated with Concept Country or Alibaba Cloud.
| Build | Size | Same next token as bf16 |
|---|---|---|
| netprojects/Qwen-Concept-27B-MLX-6bit (recommended) | 22.8 GB | 98.6% |
| netprojects/Qwen-Concept-27B-MLX-8bit | 29.5 GB | 99.0% |
| netprojects/Qwen-Concept-27B-MLX-4bit | 16.1 GB | 95.3% |
| netprojects/Qwen-Concept-27B-MLX-bf16 | 54.7 GB | reference |
The shorter thinking holds after quantisation
Hard math problems (MATH-500, level 5) with thinking on and greedy decoding, the same problems and settings for every model, on a MacBook Pro M5 Max (128 GB), 10 November 2026.
40 problems, 3,072-token cap, both models at 4-bit:
| Plain Qwen3.8-27B 4-bit | Qwen Concept 27B 4-bit | |
|---|---|---|
| Thinking tokens, mean | 1,746 | 1,202 (31% fewer) |
| Thinking tokens, median | 1,477 | 673 |
| Finished thinking within the cap | 58% | 83% |
| Correct final answer | 55% | 73% |
| Time for the run | 19 min | 12 min |
The first 16 of those problems with a 2,048-token cap, mean thinking tokens: plain Qwen3.8-27B 4-bit 1,178; Qwen Concept 4-bit 876, 6-bit 801, 8-bit 909. Concept Country reports 27 to 54% fewer thinking tokens on its own benchmarks at Q8, so the effect carries over to every build here. The plain model is mlx-community/Qwen3.8-27B-4bit. Answers are scored by exact match of the boxed result; most of the accuracy difference comes from finishing within the cap more often.
"Same next token" in the table above compares each build with the MLX bf16 build over 12 prompts (thinking on and off), 96 teacher-forced tokens each.
Use
pip install -U mlx-lm # text
mlx_lm.generate --model netprojects/Qwen-Concept-27B-MLX-6bit --prompt "How many primes are there below 100?" --max-tokens 4096
pip install -U mlx-vlm # text and images
mlx_vlm.generate --model netprojects/Qwen-Concept-27B-MLX-6bit --image photo.jpg --prompt "Describe the scene." --max-tokens 1024
mlx_lm.server and mlx_vlm.server serve it with an OpenAI-compatible API. Tested with mlx 0.32.4, mlx-lm 0.32.2 and
mlx-vlm 0.7.11. For thinking mode the Qwen3.8 card recommends temperature 1.0, top-p 0.95, top-k 20.
What was checked
- The edit is intact. Compared tensor by tensor with plain Qwen3.8-27B, 842 of 851 language-model tensors are
identical and exactly nine differ:
mlp.down_projof layers 32, 34, ..., 48, each by 0.4 to 0.6% in norm. This matches theEDIT_SPEC.jsonthat ships with the model and is included here. - Source files. All 33 files of the upstream bf16 folder match its
SHA256SUMS. - Images. The 4-bit, 6-bit and 8-bit builds each describe a test image correctly in mlx-vlm.
License
Apache-2.0, as the original model and Qwen3.8-27B. LICENSE is the Apache-2.0 text from the upstream repository and
NOTICE describes the conversion and credits Concept Country and Alibaba Cloud.