v1.0
Initial release — ~8M parameters, 1,024-token context, a real GGUF export. Runs in LM Studio, Ollama, llama.cpp, and the browser.
What it's like to talk to it
PorchLM v1 talks like a small, opinionated autocomplete more than an assistant. Its default voice leans 18th- and 19th-century — the training corpus is public-domain text (Wealth of Nations, Plato, Emerson, Aristotle, Pride & Prejudice) — and it shows in the phrasing.
Replies are short and sometimes open with a brief <think> block that names the topic before answering — a learned habit, not reasoning. It has no memory outside the current conversation, no internet access, and gets facts wrong often, because it was never given many to begin with.
What it was trained on
PorchLM v1 was trained on simple books like Percy Jackson, Moby Dick, and similar. This made it prone to acting out characters. A specific incident occurs quite often where it discusses that it's sister is getting married and that it hasn't talked to her in a while.