Qwen3.8-27B Does STM32 Design

For an embedded electronics designer, the frontier AI model is a meter running in the background — every register lookup, every code review, every question ticks the cost odometer slowly over. This episode is about building a fully LOCAL AI that does real, daily, professional STM32 microcontroller design work at zero token cost — offline, private, and on a surprisingly modest machine. The PC: an AMD Ryzen 5 5500 desktop with 64GB RAM and two RTX 3060 graphics cards, running a Qwen3.8-27B model at 32K context. It generates at 17 tokens per second, which is roughly three times faster than a human can read. SEVEN MCP servers give Qwen3.8-27B-Q4_K_M superpowers: the Forth gateway, so the AI talks to the real chip over a debug probe the Mecrisp knowledge server, which knows the exact Forth dialect the Regmon live-register monitor, reading the silicon's actual state the Saleae logic-analyser server, decoding the real waveforms LiveCheck, bridging the model to the editor's live state the Fossil SCM server, giving it the project's tickets, wiki and history the Xschem schematic server, so it can read the actual circuit diagram Together they make the model genuinely grounded — answering real design questions like "which GPIO pins feed the STM32F051 comparators in window mode, and can the output drive a counter?" with datasheet accuracy, because the facts come from the manufacturer's own SVD database and the live chip. Zero hallucinations, STMicro ground truth hallucenationsonly. This rivals the commercial STM32 ecosystem without the drama, the LSP architecture means you keep your favourite editor, and the honest economics — the frontier model's cost is front-loaded to build all this tooling, and then the daily work runs at zero marginal cost, forever. This is the longer-form episode for electronics designers who've wondered whether a local model could ever be serious about real work. The answer is yes — but not because of the model alone. Because of what you build around it. Produced with the help of AI without which this video would not have been possible