Vinci

Vinci is an open-weight, 4-billion-parameter chat model from independent developer George Pu. The pitch is narrow and deliberate: small enough to run on a consumer laptop, with two choices baked into post-training — answers tuned toward directness rather than agreement, and a defined personality instead of the flat assistant register most compact models ship with. The weights are published, so you download and run it yourself rather than calling a hosted endpoint.

It is aimed at people who want a model they own outright: privacy-conscious users keeping conversations off third-party servers, hobbyists running inference on their own hardware, and developers who want a cheap base to fine-tune or embed. Four billion parameters puts it in the most crowded tier in open weights, where the competition has sharpened considerably — the major labs now ship compact variants of their flagship families, and the Chinese open-weight releases have pushed capability-per-parameter hard. Vinci is fast and memory-light. It is not competing on reasoning, long context, or code with anything hosted, and that is not a criticism so much as a description of the weight class.

The bet is on character rather than benchmark position, and that is still the honest read some weeks past launch. Most small open releases optimize for eval scores; Vinci's framing leans on honesty and voice instead. What would settle whether the bet paid off — published benchmarks, an independent evaluation, a visible base of people running it in earnest — was not part of the launch, and we have not been able to confirm any since. Until something surfaces, the differentiator is a claim about the tuning rather than a measured property, and the project carries the usual solo-maintainer risk around updates, support, and whether it gets a second release at all.