本文来源:https://mp.weixin.qq.com/s/GvMIaiBSTzHRrFc0gSVU7Q
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。



Source: https://mp.weixin.qq.com/s/GvMIaiBSTzHRrFc0gSVU7Q If you have any copyright questions, please write to: 13951431913#139.com or submit a comment on the right.
If you happen to need to install a reasonably capable local model for high-privacy work,
and you happen to be using a performing computer, like a mac with 32GB of RAM,
and you happen to be using WorkBuddy,
then the local model built into WorkBuddy is worth a try.
In "Settings" — "Model", you can see a "Local Model" added to the model list, which can be downloaded and used in WorkBuddy with one click, without needing to use ollama or llama.cpp to host it.
This Qwen3.6 35B-3AB is an MoE model, 35B parameters activating only 3B parameters, improving local inference speed. At the same time, Tencent performed pruning optimization on the model, making it only 15GB in size, 1/3 less than the 24GB version on ollama, which makes it run smoother in WorkBuddy.
The model speed feels decent in practical use, with a stable generation rate of 30–35 tokens/s, and a memory usage of about 17GB during operation. The first character latency is relatively high, mainly because the WorkBuddy platform prompt and tool list loading are very large (often reaching tens of thousands of tokens), turning off unused tools can improve this.
Compared to the cloud, it is certainly not as fast, and the context is not as stable as online models; but as a local, non-internet, zero-metering solution, its speed is already usable for daily tasks, offering excellent cost-performance.