Two Free AIs Just Dropped: One Codes Faster, One Knows Canada

A small team spent a few weeks with a rented high-end Dell server and used it to build two free AI models, posted on Hugging Face for anyone to download. These are "open-weights" models, meaning you can run them on your own machine instead of paying a monthly subscription. One is built for coding, the other for Canadian questions.

The coding model, Victoria, is a slimmed-down version of an existing AI called Qwen Flash Next. The team removed 44% of its "experts" — specialist chunks the AI picks from when answering — making it smaller without much loss in quality. It zips through 280 words a second on one server chip, about twice as fast as before, and uses 35% fewer words to get the same job done. It scores 70% on a tough test where AI must complete real computer tasks end to end.

Maple tackles a quieter problem: most AI models assume you live in the United States. Ask about taxes or benefits and you get American answers. Maple is tuned to default to Canada. On 600 test questions, it cited an official Canadian source 63% of the time, up from just 6%. Fully correct answers rose from 6.6% to 22%, and it stopped saying "I don't know" nearly as often.

The catch: the results were graded by a panel of AI judges, not humans, so treat the numbers as promising but unproven. There is also a download snag — a popular free tool called llama.cpp does not yet recognize one file, so you need a special build to run it. And both models still need a powerful computer and some technical patience; this is not a plug-and-play app.