What Span-02 is
Span-02 is Respan’s general decision model, released on 15 October 2026. It is a different model from Span-01, Respan’s closed-weights classifier for behaviors in agent traces, with its own endpoint, request format and price. Respan’s model card credits Harvey Li as the author.
You send a state, text or any JSON value plus up to four images, and typed questions: Choice, Score, Noul, or a number question that returns a value with a confidence. Each answer carries its full probability distribution.
How it works
Span-02 reads the option labels’ probabilities in one forward pass per question, with the shared state computed once. When the top answer is below 0.8, it can write a short thought, read again and average the two results. You control this per request: auto thinks only when unsure, on always thinks, off never does, and budget_ms sets a deadline. Respan calls this anytime reasoning, because the model answers with its best decision so far when the deadline arrives.
It is a fine-tune of Google’s Gemma 4 26B-A4B, a mixture-of-experts model with about 4B parameters active per token.
Access and price
Respan’s API takes POST https://api.respan.ai/api/v1/decisions with a body in the System One format, so a client written for Jev works with a changed URL. It costs $0.01 per million input tokens with free output, on Respan and on OpenRouter. The weights and Respan’s own Rust and CUDA serving engine are on Hugging Face under Apache 2.0, with the weights also under Google’s Gemma 4 terms. They run only with that engine, on an NVIDIA Blackwell GPU.
What the benchmarks show
On the Jev Decision Index 0.4.1, generated on 22 October 2026, Span-02 scores 59.07 on the Full score, 49th of 229 rows, against Jev 1.14’s 60.24. Its public score is 63.62, from a run its author submitted. Its median latency there is 33 ms. Respan’s card, last updated on 16 October, still shows 62.6 and a tied fourth place, a result from before the 0.4.1 file.
Its calibration error on the Index, the average gap between stated confidence and how often it is right, is 0.114, against Jev 1.14’s 0.0799. Sorted from lowest, it is 172nd of 229 rows, so a confidence threshold tuned on another model may not carry over.
On Benchmark Heaven’s JevBench, measured on 14 October, it has a Capability Score of 69.3, from Intelligence 60.9 and Calibration 77.8.
Specifications
| System One Score | 82 from 2 of 3 boardsJev Decision Index 82.6, JevBench 81.3. How the score works |
| Question types | ChoiceScoreNoul |
| Max Choice options | 679 |
| Score levels | Up to 10 |
| Questions per call | Not documented |
| Total context | 131,072 tokens |
| State budget | Not documented |
| Rate limit | Respan's docs give Span-02 its own per-organization rate limit, separate from chat completions, and return 429 above it. The size of the limit is not published. |
| Endpoint | POST https://api.respan.ai/api/v1/decisions |
| Licence | Apache-2.0 |
| SDKs |
The model card gives a 131,072-token context per question, up to 4 images (PNG or JPEG, 16 MiB each) and a 64 MB request body. Choice takes 1 to 679 options and Score 2 to 10 levels. A number question returns the number itself. The API does not stream. Self-hosted, the weights run only with the included Rust and CUDA engine on an NVIDIA Blackwell GPU (RTX PRO 6000 or RTX 5090) and need about 28 GB of GPU memory.
Versions
- span-02, 15 Oct 2026, Model name span-02 or respan/span-02 on Respan's API. OpenRouter lists it as respan/span-02, snapshot span-02-20261015. Weights on Hugging Face as respanai/span-02, first published as lihe07/span-02, the account the Jev Decision Index and JevBench rows link to. A fine-tune of Gemma 4 26B-A4B (about 4B active parameters), stored as NVFP4 experts with FP8 dense layers, 16.4 GB. Release notes
Use cases
What people use Span-02 for, one page per pattern.
Support inbox triage with System One models
Send a support ticket to Jev once with every question attached. Category comes back as a selected label, severity and frustration as numbers on scales you wrote, refund intent as a probability. Your code reads those values and decides what happens to the ticket.
Intent and model routing with System One models
One Jev call reads an incoming request and returns its intent as a label plus a difficulty rating on a scale you wrote. Your router reads both numbers and picks the handler: deterministic code, a cheap model, an expensive one, or a human queue.
Confidence-gated actions with System One models
Jev returns a confidence value from 0 to 1 alongside every Choice and Score answer. Your code treats it as a separate axis: act automatically when it's high, confirm or flag when it's middling, hand the decision to a person when it's low. Riskier actions get higher bars.
Agent routing and skill selection with System One models
An agent choosing from a long skill roster reads one truncated line per entry and often loads the wrong thing. Jev ranks every entry in one request and separately answers whether any skill applies at all, so the agent gets a short hint instead of a guess.
Benchmarks
Published scores, each linked to its source. A row marked vendor-run was measured by the vendor itself, so read it as one side's result. The compare hub sets these beside the other models.
| Benchmark | Metric | Score | Run by | Date |
|---|---|---|---|---|
| Jev Decision Index 0.4.1 | Full score | 59.07 score | Independent | 22 Oct 2026 |
| Jev Decision Index 0.4.1 | Public score | 63.62 score | vendor-run | 22 Oct 2026 |
| Jev Decision Index 0.4.1 | Expected calibration error | 0.114 ECE | Independent | 22 Oct 2026 |
| JevBench 1.7.47 | Capability Score | 69.3 score | Independent | 14 Oct 2026 |
- Jev Decision Index, Full score: Row: span-02 (Gemma 4 26B-A4B), 49th of 229 rows by this score. Counts 60% private same-skill tests and 40% private new-domain tests; public benchmarks no longer count. 0 is random guessing, 100 is perfect. The Index ran the private tests on one RTX PRO 6000 with Respan's engine (NVFP4 experts, FP8 dense). The row notes gated short thinking on about 19% of questions, and that serving settings were compared on Decision Index public samples (author-disclosed). Jev 1.14 scores 60.24.
- Jev Decision Index, Public score: Row: span-02 (Gemma 4 26B-A4B). The author submitted the public run, and the Index rescored it with its own kit. Jev 1.14 scores 60.7.
- Jev Decision Index, Expected calibration error: Row: span-02 (Gemma 4 26B-A4B). Lower is better. Jev 1.14 scores 0.0799.
- JevBench, Capability Score: Row: Respan Span-02, self-hosted, measured 2026-10-14. Intelligence 60.9, Calibration 77.8. Composite 51.5, 39th on the open board, above the board's official cost cap.