🪨 Quartz Micro Preview
Chat model: Quartz Micro Preview V3 (instruction-tuned, 100M parameters)
The newest Quartz: the same 100M-parameter size as V2, but pretrained on about 65× more text (~6.5B tokens) and tuned on 44k conversations that include identity data, grounded Q&A and "I can't do that" examples. In our head-to-head it beat V2 on 25 of 33 judged prompts and knows who it is (71/100 held-out identity checks vs 3/100 for V2). It is still a tiny model: it states wrong facts and invents details, so don't rely on it for anything that matters. It sees the earlier messages of the chat (up to its 1,024-token window). Ask "what are your limitations?" for a longer straight answer (that reply is fixed text, not generated). Questions about who it is are answered by the model itself.
Ask something, and keep your expectations low.
Run one prompt against any two Quartz models and read the answers side by side. Both get the same settings and the same random seed, so differences come from the models. This tab shows raw model output only (no scripted answers). Base models are not chat models: they just continue your text, so try a sentence start like The history of the Roman Empire.