I'm implementing an MCP client using Qwen3 4B and its tool call capabilities are impressive! I'm sure it will only improve and the 30B is probably already much better.

What are you running it in, ollama? Did you have to install some additional software to enable it to call tools (also via MCP?)

LM Studio. No additional software but I implemented the MCP client myself using the typescript-sdk[0]

0. https://github.com/modelcontextprotocol/typescript-sdk

Thank you very much!

I use it on Windows, I am just loosely aware that I could run it on debian if I wished. I use 7b and its roughly as useful as GPT 3.5. I dont have any tools linked to it yet.

Does that mean "pretty useful" or "a total waste of time"? I never got much useful code out of GPT 3.5.

3.5 was all about context window management. The second you hit that wall it became useless. I would either chunk the problem into smaller pieces or rollover my context every hour or so and it worked a treat.

The second it started dropping context it was hilariously unreliable. It would just, rewrite the entire function or drop elements we hadnt discussed for a while.

I see, thanks!