Respan releases Span-02, an open-weight decision model that can think up to a deadline
Respan AI released Span-02 on 15 October 2026. Its first model, Span-01, launched behind a waitlist and only judged behaviours in traces. Span-02 is a general decision model with open weights on Hugging Face, a hosted API at $0.01 per million input tokens and an OpenRouter listing.
What’s different. You can give each request a thinking deadline in milliseconds. With thinking off, Span-02 reads answers straight from option probabilities, like other decision models. In the default auto mode it writes a short hidden thought only when its top answer is below 0.8, then reads the answer again. It also adds a number answer type on top of choice, noul and score, and takes up to four images. The weights are a Gemma 4 26B-A4B fine-tune quantized to 16.4 GB and only run on Respan’s own Rust/CUDA engine, on one Blackwell GPU (RTX PRO 6000 or RTX 5090).
Numbers (UnverifiedClaim). The card reports a full run of Hugging Face’s Decision Index 0.3 public suite: 63.62 on text and 70.69 on vision. Respan ran it; Span-02 isn’t on the official board yet. The blog shows a 12–0 chess match against Jev 1.13 and a coordinates-only land map where accuracy climbs from 71.9% with thinking off to 84.7% at one second. OpenRouter’s own measurement is more sobering: 0.24 s p50 but 5.8 s p99 end to end.
On this site. Span-02 gets its own page at 6.6 maturity / 7.0 capability / 22 adoption. Open weights plus a paid API and OpenRouter lift maturity well above Span-01 (3.5). Capability stays at 7.0 because the benchmark is self-run and calibration has no published figure.