Cicero

Butler Compute's decision model (released Nov 5): a Qwen3.5-35B-A3B post-train (3B active) answering noul / choice / score over text, a chat, or up to four images, 131K context, up to 64 questions per call — letter-logit readout, one forward pass, per-type calibration. Served on Butler's own API at $0.01/M input with output free (level with Span-02 as the cheapest commercial input price), on both the System One and OpenAI Decisions wire shapes, with Apache 2.0 weights and an NVFP4 build published. Its JDI 0.4.1 public-suite run (71.33, vendor-run) has no board row yet — the card is upfront that until an independent board scores it, the numbers are Butler's.