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

13 hours ago

Qwen3.5 models are effective for local agentic coding sessions, even on older hardware like a 6-year-old MacBook M1.
OpenCode and LM Studio were used to create a Telegram bot that forwards messages to a local LM Studio OpenAI-compatible server.
The setup involves a larger model (Qwen3.5 9B) for coding and a smaller model (Qwen3.5 0.8B) for chat responses.
Installation steps include setting up OpenCode, configuring it to use LM Studio locally, and installing the necessary models.
A minimal Telegram bot example was provided, showcasing how to forward messages to the local LLM and return responses.
Performance is slow on an M1 MacBook but usable for small, sensitive, offline tasks.
The setup is practical for teams needing privacy and offline capabilities, though it doesn't replace higher-end coding stacks yet.
Future tests on newer hardware like M4/M5 or stronger desktops could show improved performance.

Hasty Briefsbeta