Skip to main content
Introducing Span-02
Release · Oct 15, 2026

Span-02Anytime reasoning for decisions

Multimodal Numerical outputs Open weights

Try the playground$0.01/M input tokens

Stop reasoning anytime.

Give it a deadline, and it returns its best decision so far.

Thinking off is frontier-level at the lowest latency. Set budget_ms to think longer, or let auto mode's learned gate choose.

Decision Index 0.3, public
56575859606162636430 ms100 ms300 ms1 sLatency per decisionDecision IndexPerplexity DeciderCloudflare clefJev 1.13.0Fastino GLiDESurogate Rune 26BvLLM-SR d3-flashXor 26B (Juspay)Bespoke Nimble 9BCloudflare clef-flashSpan-02 (thinking off)Span-02 (thinking 1 s)
Decision Index 0.3, public index: 37 benchmarks, chance-corrected (0 is guessing, 100 is perfect).

Playing chess with Jev

12–0 against Jev, even with Span-02's reasoning turned off.

One of the 12 games: Jev 1.13 (white) against Span-02 (black, reasoning off), from a set opening. Span-02 mates on move 11. Beside the board, each side's top candidate moves and think time per move. Played on Oct 15, 2026.

Browser use

We gave Span-02 a browser-use harness. In a Wikipedia race, it finished nearly 2× faster than GPT-6 Luna: 20.6 s vs 38.5 s. The race: get from one article to another by clicking only links in the text, so every click involves reasoning.

Alexander the Great → University of Milan, clicking only links in the article, shown at 2.5×. Each pick is boxed on the frame it was made on; Span-02's runners-up show their probability. GPT-6 Luna's goal included the hint "Italy". Both runs on Oct 15, 2026.

Numbers, not just yes or no.

Span-02 outputs numeric values directly, with how sure it is.

A four-panel comic: a student at the board writes 7 + 7 =, looks back at the class, a classmate holds up a sign reading 13: 0.0, 14: 1.0, 15: 0.0, and the student copies those probabilities onto the board.
jev@JinjingLiang ↗
The same comic for Span-02: the classmate's sign reads 14, more than 99.9% sure, and the student writes Span-02's answer on the board as JSON: number 14, confidence 0.9999.
span-02our edit
span-02 / numbers
700 / 400 =
NumberWhat is the calculated answer?
1.75
{ "number": 1.75, "confidence": 0.9996 }thinking off

Playground

span-02 / playgroundlive
Image
Questions
Number
Yes / no
no limit

answer

NumberHow many people are here?
—
Yes / noIs the background blue?
—

Press Run to see the decision.

Blind Earth

No map, just coordinates. For each of 4,050 points on Earth, Span-02 answered “is this point in water?” from latitude and longitude alone. The longer it thinks, the sharper the map.

The answer key: land and sea from Natural Earth on a 4° grid.
Answer keyNatural Earth
12 Claude modelstheir scoring

Brightness is P(land) on a 4° grid (4,050 points, one per request), scored against Natural Earth. Jev 1.13.0 scores 79.6% on the same question. South America is still faint. The idea and the Claude chart are from @celestepoasts on X; that chart scores against a 1-km land mask, weighted by area, so its percentages don't compare with ours.

span-02, by thinking time
Span-02 with thinking off: broad smudges of land across the north.
Thinking off71.9%
Span-02 with 100 ms of thinking: much like thinking off.
100 ms72.5%
Span-02 with 300 ms of thinking: North America, Eurasia, Africa and Australia take shape; South America stays faint.
300 ms84.5%
Span-02 with 1 s of thinking: much like 300 ms.
1 s84.7%

Start using Span-02

On the Respan Gateway it costs $0.01 per million input tokens, and output is free: the cheapest paid decision model there. Or run the open weights.

span-02 / get started
apiSystem One endpoint compatible
curl https://api.respan.ai/api/v1/decisions \  -H "Authorization: Bearer $RESPAN_API_KEY" \  -H "Content-Type: application/json" \  -d '{    "model": "span-02",    "state": "Customer: I was charged twice this month.",    "questions": { "team": { "type": "choice",      "instructions": "Which team should handle this?",      "criteria": { "billing": "Billing", "tech": "Tech support" } } },    "reasoning": { "mode": "on", "budget_ms": 300 }  }'
Docs ↗Playground ↑
open weights

Host it yourself.

  • 1M tokens/s on a single RTX 5090
  • Up to 7× faster inference than vLLM
Weights on Hugging Face ↗
Frank Chen
Frank Chen@francchen

Built something with Span-02?
We'll repost it.

Quote on X ↗