Benchmarks & research
Maincode's open-weight Matilda Jev answers yes/no, choice and score questions about text, JSON or images, returning a probability for every option.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
Despite the name, it is Maincode's own model, not TypeSafe's: a Qwen3.8-27B backbone with a 255-option readout, released under Apache-2.0 with a bundled server that accepts Jev-format requests. The bf16 weights need about 49 GiB, and Maincode has only validated this build on AMD MI355X accelerators.
Matilda Jev picks between answers you supply and says how likely each one is, such as whether a customer's message is about card delivery, a double charge or a cash withdrawal. It runs on your own equipment but needs a very large graphics card. Maincode's test score comes from its own run and has not yet been checked by the people who run that test.