Mantus 3 is here
The default Ichnus model gets a new version — and the release notes are in plain language.
Mantus 3 is now the default model across Ichnus. For most people, the change is invisible, which is how a default should feel. For everyone else, here are the release notes without the jargon.
Better at long work. Mantus 3 continues the work Mantus 2 started on long documents — it is more reliable at finding and combining the parts of a long text that answer your question, and less likely to lose track of a conclusion while it reads. If you give it a big messy document, it does better.
More honest about not knowing. The calibration work from our deference-under-uncertainty line is in this release. Mantus 3 is less likely to sound certain when it is guessing. It still guesses — every model does — but it is better at flagging when it does.
Faster to start. First-token latency is down, mostly from the serving work that made Krus 2 faster and is now benefiting the bigger models too.
The safety story is unchanged and that is the point. Mantus 3 passed the same evaluation suite as every release before it, reviewed by the Safety & Evaluations Committee. The full evaluation report — what we tested, what we found, what we did not test — is published alongside the model. We did not write “safer” in this post because the process, not the word, is the claim.
Mantus 3 is free, like everything else. If you find it doing something wrong, tell us — the report is how we know, and you are part of it.
All news →