I'm interested in using AI for free, but recently the free quota for Codex has been reduced
and the generous Gemini CLI, which only gives 20 minutes of availability per day, has become disappointing. In the meantime,
I discovered a usable local model called Qwen3.6 35B-A3B, loaded it using the Ollama API, installed the Qwen CLI, and just successfully got it running.
It seems there's no major difficulty in reading files, making decisions, and writing, as I'm familiar with using the tool.
Now I've secured one AI slave that I can use comfortably. I think I'll need to use it more to know the details.
And here are the performance test metrics.
The equipment is AMD Ryzen 5600, RAM 64GB, RTX 3090 x1, and I'm using the model from the title via Ollama API.
In summary, the performance is approximately 17 tokens per second.
Below are the results.
The measurement method was to run # ollama run qwen3.6:35b-a3b --verbose and send a test prompt
total duration: 6m22.707600455s
load duration: 172.211636ms
prompt eval count: 173 token(s)
prompt eval duration: 459.077036ms
prompt eval rate: 376.84 tokens/s
eval count: 6343 token(s)
eval duration: 6m18.318385948s
eval rate: 16.77 tokens/s
and obtained results in this manner.
Other information.
# ollama ps
NAME ID SIZE PROCESSOR CONTEXT UNTIL
qwen3.6:35b-a3b {ID value} 31 GB 23%/77% CPU/GPU 131072 Forever
That concludes this post.
