2x Qwen 3.5 on M1 Mac: 9B builds a bot, 0.8B runs it

advanced-stack.com

I was trying this the other day with opencode and ollama and it all seemed kind of broken/useless.

I don't understand the two llm approach plus telegram here?

It seems like the bot is both creating the telegram interface and using it to be a coding assistant?

It's mostly a test of the coding capabilities of the 9B model. The 0.8B is used to make the telegram bot smarter than a if/then/else.

I find LM Studio more usable for local setups (desktop/laptop) and I would use directly llama.cpp stack for a (local) server deployment

I revisited this and using the models directly via ollama run was actually surprisingly fast.

A bug or misconfiguration with the connection to opencode seemed to be the culprit.

mmh.. interesting; I'll try to compare the perf