Concept Country's Qwen Concept 27B on a Mac: less thinking, and it survives quantisation

NetProjects13 Nov 2026

Reasoning models think before they answer, and on hard problems that thinking can run to thousands of tokens. On a laptop every one of those tokens costs time, and if the thinking runs out of room before it finishes, there is no answer at all. A model that reaches the same answer with less thinking is faster and finishes more often.

Concept Country's Qwen Concept 27B is Qwen3.8-27B with nine weight matrices rewritten so that it does exactly that. It is released with open weights under the Apache-2.0 licence. We ported it to MLX so it runs on Macs with Apple Silicon, with text and images, and we checked the one thing Concept Country's own card leaves open: whether the shorter thinking survives quantisation. It does, in every build we published on Hugging Face.

Always up to date. We keep these builds current on Hugging Face. When Concept Country releases a new version of Qwen Concept 27B, 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.

Hugging Face: https://huggingface.co/netprojects
NetProjects MLX builds collection: https://huggingface.co/collections/netprojects/netprojects-mlx-builds

What the model does

Concept Country calls the method weight-space programming. There are no gradient steps and no training data: the edit is derived from the model's own activations and compiled into nine MLP down-projections. Against the unmodified Qwen3.8-27B, Concept Country reports GPQA Diamond up 4.3 points, LiveCodeBench v6 up 13.0 points and 27 to 54% fewer thinking tokens, measured with its bf16 weights served at Q8_0.

Concept Country also ran a control: a random edit of the same layers and the same size, in a different direction, does nothing measurable. That is what shows the direction of the edit matters, not just the fact that the weights were changed. The paper, data and code are at https://concept.country/research/weight-space-programming and Concept Country's original is at https://huggingface.co/ConceptCountry/qwen-concept-27b. Concept Country describes the model as experimental and under evaluation.

What we released

Four MLX builds on Hugging Face, converted from Concept Country's bf16 weights at revision 4c96373. The quantised builds quantise the language model (affine, group 64) and keep the vision tower in bf16, so every build runs in mlx-lm for text and mlx-vlm for text and images.

Hugging Face	Size
netprojects/Qwen-Concept-27B-MLX-6bit (recommended)	22.8 GB
netprojects/Qwen-Concept-27B-MLX-8bit	29.5 GB
netprojects/Qwen-Concept-27B-MLX-4bit	16.1 GB
netprojects/Qwen-Concept-27B-MLX-bf16	54.7 GB

We tested the builds with mlx 0.32.4, mlx-lm 0.32.2 and mlx-vlm 0.7.11.

How well it matches the original

We first checked that all 33 files of Concept Country's bf16 folder match its published SHA256SUMS. Then we compared the weights tensor by tensor with plain Qwen3.8-27B: 842 of the 851 language-model tensors are identical and exactly nine differ, the MLP down-projections of layers 32, 34 and so on up to 48, each by 0.4 to 0.6% in norm. That matches the edit specification that ships with the model, which we include in every build.

Next we compared each quantised build with the bf16 MLX build, the full-precision weights, over 12 prompts with thinking on and off, 96 teacher-forced tokens each, counting how often the build predicts the same next token.

Measured on Qwen Concept 27B, 10 November 2026, MacBook Pro M5 Max (128 GB).

Build	Same next token as bf16
MLX bf16	Reference
MLX 8-bit	99.0%
MLX 6-bit	98.6%
MLX 4-bit	95.3%

The 4-bit, 6-bit and 8-bit builds each describe a test image correctly in mlx-vlm.

The shorter thinking holds after quantisation

Concept Country's card says its quantised files are deployment conveniences and that nothing there claims a 4-bit conversion preserves the effect. So we tested it. We gave plain Qwen3.8-27B and Qwen Concept 27B the same hard maths problems, MATH-500 level 5, with thinking on, greedy decoding and the same settings for both. The plain model is mlx-community/Qwen3.8-27B-4bit, so both models ran at 4-bit.

Measured on Qwen Concept 27B, 10 November 2026, MacBook Pro M5 Max (128 GB), 40 problems, 3,072-token cap.

	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

Answers are scored by exact match of the boxed result. Most of the accuracy difference comes from finishing within the cap more often: with less thinking to do, Qwen Concept gets to an answer before it runs out of room.

Measured on Qwen Concept 27B, 10 November 2026, MacBook Pro M5 Max (128 GB), the first 16 of those problems, 2,048-token cap.

Model	Thinking tokens, mean
Plain Qwen3.8-27B 4-bit	1,178
Qwen Concept 27B 4-bit	876
Qwen Concept 27B 6-bit	801
Qwen Concept 27B 8-bit	909

Every build thinks less than the plain model, so the effect Concept Country reports at Q8_0 carries over to all of our builds. Our figures use our own problems and settings and are not a reproduction of Concept Country's benchmarks.

Try it

On a Mac with Apple Silicon, with mlx-lm for text:

pip install -U mlx-lm
mlx_lm.generate --model netprojects/Qwen-Concept-27B-MLX-6bit --prompt "How many primes are there below 100?" --max-tokens 4096

With mlx-vlm for text and images:

pip install -U mlx-vlm
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. For thinking mode, the Qwen3.8 card recommends temperature 1.0, top-p 0.95 and top-k 20. The model cards have the full details:

MLX 6-bit: https://huggingface.co/netprojects/Qwen-Concept-27B-MLX-6bit
All four builds: https://huggingface.co/collections/netprojects/netprojects-mlx-builds

Honest limits

All figures come from one machine.
Our thinking test is 40 hard maths problems with greedy decoding and a token cap, not Concept Country's benchmarks.
We have not run GPQA Diamond or LiveCodeBench ourselves.
We have no speed figures beyond the run times above, and no memory figures.
We checked images only by asking each quantised build to describe one test image.
Concept Country describes the model as experimental and under evaluation.
The 4-bit build differs more from the original than the others.

Licence

Qwen Concept 27B, Qwen3.8-27B and our builds are all under the Apache-2.0 licence. Each of our repositories carries the Apache-2.0 text from the upstream repository and a notice that describes the conversion and credits Concept Country and Alibaba Cloud.

Credits

Qwen Concept 27B is by Concept Country, built on Qwen3.8-27B by Alibaba Cloud's Qwen team, under the Apache-2.0 licence. The MLX builds are by NetProjects (Internet Projects Ltd). Not affiliated with Concept Country 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.