Original video in Chinese.
Key Takeaway
- Qwen 3.8 27B has become a usable productivity model: even a quantized local version can reliably drive multi-agent orchestration (calling Skills, analyzing requirements, assigning tasks to sub-agents). It marks the point where compact frontier models are truly entering productivity scenarios, with markedly improved capability per parameter.
- The barrier to local deployment has been greatly lowered: it can run on consumer-grade devices. 24GB on a Mac is barely usable, 32GB is the practical starting point, and 48GB can support longer contexts. I’m planning to switch to a Mac Studio to get higher memory bandwidth and longer context, so I can take local computing power into my own hands for everyday tasks.
- Model stratification and dynamic routing are the direction of the future: 3B-8B handles simple tasks, 14B-35B handles local content and agent execution, 70B+ handles complex reasoning, and trillion-parameter models handle high-value research. This layered approach is the sensible path for AI adoption, rather than sending every task to a giant cloud model.
After using Qwen 3.8 27B, I suddenly wanted to buy a Mac Studio. Because I want larger unified memory and higher memory bandwidth, so local models can run faster while also fitting longer contexts. I want to keep part of my computing power in my own hands and be more free.
My current MacBook Pro only has 48G of memory. Running the 4-bit, MLX version of Qwen 3.8 27B gives me a speed of about 14. Slow as it is, it can run. If I use the non-MLX version, the speed is only about 10, which is unacceptable.
The quantized version running locally makes some compromises in performance. To minimize the impact of local quantization as much as possible, I also tested the cloud high-precision version through OpenRouter.
I connected Qwen 3.8 27B to my own newtype OS and changed the model for each Agent to this one. In actual testing, I found that this model can fully drive a multi-agent orchestration framework—it can call Skills, analyze requirements, assign tasks to sub-agents, and so on.
Honestly, I was pretty surprised.
You have to understand, this is a 27B model, not one of those giants. Even if you use the quantized version, the performance sacrifice isn’t much. But the deployment barrier is much lower, and consumer-grade GPUs can run it.
A Mac can barely run it with 24GB, 32GB is the more practical starting point, and 48GB can give you a larger context. If you move up to an M4 Max and 128GB of unified memory, not only will the speed be much faster, but there will also be more room for context, and you can even run multiple models at the same time.
That’s why I want to buy a Mac Studio. As I said in Knowledge Planet: run it locally and supply it to agents like Claude Code and OpenCode, where it will be dedicated to everyday tasks. For example, document processing, information gathering, and so on. Low cost, good privacy, and no risk of being controlled by others.
I believe Qwen 3.8 27B will definitely become a milestone. It marks the beginning of compact frontier models becoming usable productivity models. A 27B model proves that the parameter scale required to reach frontier capability is dropping rapidly. Behind this round of progress is the fact that the effective capability carried per parameter has increased.
If we give it another year, we will very likely see model stratification and dynamic routing enter large-scale use.
Models from 3B to 8B will handle classification, extraction, rewriting, and simple tool calls.
Models from 14B to 35B will handle most local content production and Agent execution.
Models from 70B to 120B will handle complex reasoning and high-value tasks.
And trillion-parameter models will handle research, planning, judgment, and so on.
This kind of stratification and routing is the right path for further AI adoption. Otherwise, if everyone and every task has to call a giant cloud model, that just doesn’t make sense economically or energetically.
Finally, I recommend everyone try Unsloth Desktop. Running models locally is very convenient and very clean. It scans and displays all the models downloaded on your machine, including those downloaded from Ollama and LM Studio. You can also search for whatever model you want to download through it. If you want to supply it to agents like Claude Code, you only need to run one command.
OK, that’s all for this episode. If you want to learn about AI, want to reclaim individual sovereignty, and want to find like-minded people, come join our newtype community. See you next time!