Span-02
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity6.6/10
Open weights on Hugging Face with Respan's own Rust/CUDA engine and a Dockerfile, plus a paid hosted API ($0.01/M input) and an OpenRouter listing. TypeSafe-compatible /v1/systemone. Company-run, no SLA.
- Capability7.4/10
Choice / Noul / Score plus a number type, text and up to 4 images, optional time-boxed thinking. Official Decision Index 0.3.2: Full 62.56 (#11, Jev 60.11); JevBench open weights Intelligence 60.9, calibration 77.8. OpenRouter p50 0.24 s but p99 5.8 s. No ECE.
- Adoption22/100
Launch post at ~335 likes / ~205k views; OpenRouter shows ~971M prompt tokens in its first days. HF repo at 10 likes. No named production users yet.
Vendor claims
- Decision Index 0.3 public suite: text index 63.62 (all 140,178 requests answered), vision 70.69 (all 12,210 answered)[Vendor claim — not independently verified]
- With reasoning off, beats Jev at a quarter of the price and runs at a third of the latency of decision models at its level[Vendor claim — not independently verified]
- 12–0 at chess against Jev 1.13 with reasoning off[Vendor claim — not independently verified]
- Land-or-water from coordinates alone: 71.9% with thinking off, 84.5% at 300 ms, 84.7% at 1 s, against 79.6% for Jev 1.13.0[Vendor claim — not independently verified]
Span-02 is the second decision model from Respan AI, released on 15 October 2026. Unlike Span-01, which launched behind a waitlist and only judged behaviours in traces, Span-02 is a general System One model with open weights, a hosted API and an OpenRouter listing.
The headline feature is what Respan calls anytime reasoning. You give the request a time budget in milliseconds, and the model returns its best answer by that deadline. With thinking off, it reads the answer straight from the option probabilities in one pass, like other decision models.
Specs
| Attribute | Value |
|---|---|
| Company | Respan AI (Keywords AI, Inc.) |
| Launch | 15 Oct 2026 (X and blog, 14:00 BRT) |
| Base model | Fine-tune of Gemma 4 26B-A4B (mixture of experts, ~4B active) |
| Weights | NVFP4/FP8, 16.4 GB. Card metadata says Apache-2.0 and links to the Gemma 4 license |
| Runtime | Respan’s own Rust/CUDA engine (dllm-server, in the repo). The weights don’t load in transformers or vLLM. Runs on one RTX PRO 6000 or RTX 5090 |
| API | Local: POST /v1/systemone (TypeSafe’s System One schema). Hosted: POST https://api.respan.ai/api/v1/decisions |
| Question types | choice (up to 679 options), noul, score (2–10 levels), number |
| Input | Text or JSON state, up to 4 images |
| Reasoning | mode: auto (default; thinks only when the top answer is below 0.8), on, off. budget_ms 0–600,000 on the hosted API, 700 ms default on the local server |
| Pricing | $0.01 / M input, output free (Respan and OpenRouter); 131K context |
How it works
Each question becomes one prompt over a shared state, and the answer is read from the model’s probabilities for the option labels. If the top option is below 0.8, the model writes a short hidden thought and reads the answer again; the final distribution averages the two readings. Thoughts are capped by the time budget and decoded with Gemma’s public speculative-decoding draft. Number answers are generated token by token.
So “zero output tokens” only holds with thinking off. In auto mode an uncertain question costs a short generation. Respan doesn’t return or bill the thought tokens.
Latency
Respan says a typical request answers in tens of milliseconds. OpenRouter’s own three-day measurement shows p50 0.24 s, p95 4.1 s and p99 5.8 s end to end. The long tail probably comes from auto thinking on hard questions and from the network hop. Set mode: "off" or a small budget_ms if you need a hard ceiling.
Benchmarks (UnverifiedClaim)
The card reports a full run of the Decision Index 0.3 public suite: 63.62 on text and 70.69 on vision, with every request answered. Respan ran it itself. Span-02 is not on the official board yet (data generated 15 Oct, 168 models), and the private tasks that make up 80% of the Full score haven’t been run. The blog adds three demos against Jev 1.13 and GPT-6 Luna: a 12–0 chess match, a land-or-water map from coordinates, and a Wikipedia link race. They are fun to watch and say little about classification quality.
Official board results (checked 17 Oct 2026)
publicado (OK do Fabricio 17/10 BRT).
Two independent boards now carry Span-02. They are separate sources and are not averaged.
| Board | Result | Jev reference |
|---|---|---|
| Decision Index 0.3.2 (data of 16 Oct 19:00 UTC) | Full 62.56, #11; public 63.62 (#11); same-skills and new-domains parts feed the Full score | Jev 1.13: Full 60.11 (#18), public 57.96; Jev 1.14: 61.53 (#13) |
| JevBench open weights (draw v1.6.28) | Capability 69.3; Intelligence 60.9; Calibration 77.8; self-hosted p50 0.027 s (on-GPU, no network); cost 2.08× Jev’s reference, just outside the 2× cap | Jev 1.13: Intelligence 63.6, calibration 90.6 |
The board’s public figure (63.62) is the same number Respan reports for its own run, so the vendor claim about the public suite checks out. What the board adds is the held-out part: Full 62.56 is above Jev 1.13 and Jev 1.14 on that scale, while on JevBench Intelligence it sits slightly below Jev 1.13. The hosted API latency (p99 5.8 s on OpenRouter) is a different measurement from the board’s on-GPU 27 ms and is not replaced by it.
Fit / anti-fit
Fit when you need: an open decision model you can self-host on one Blackwell GPU, a number answer type, image input, or a dial between speed and accuracy per request.
Anti-fit when you need: standard serving (transformers, vLLM, llama.cpp), older GPUs, strict tail latency with thinking on, or independently verified accuracy.
Limits
- Custom engine only, CUDA 12.8+ and Blackwell GPUs.
- No published ECE or calibration method for Span-02.
- Benchmark numbers are Respan’s own runs.
- License: Apache-2.0 in the metadata, but the base is Gemma 4, so check the Gemma 4 terms too.
Why it is in the catalog
Span-02 has trained weights for typed decisions, published openly, plus a documented typed API that returns probabilities. That meets the catalog criteria directly. Span-01 keeps its own page as a separate, waitlisted product.
Score working
- Maturity 6.6 = availability 9 × 0.30 + docs 8 × 0.25 + integrations 5 × 0.25 + support 3 × 0.20 = 2.70 + 2.00 + 1.25 + 0.60 = 6.55. Availability 9: open weights plus a paid hosted API (same as Perplexity Decider). Integrations 5: OpenRouter in production.
- Capability 7.0
[vendor]= types 10 × 0.30 + calibration 5 × 0.25 + latency 6 × 0.25 + accuracy 6 × 0.20 = 3.00 + 1.25 + 1.50 + 1.20 = 6.95. Latency 6: p50 under 300 ms on OpenRouter, p99 well above 1 s. Accuracy 6: public suite run by the vendor, off the board (same rule as d1 and Clef). - Adoption 22: solid launch reach and real OpenRouter traffic, no named production users.
Score update (17 Oct 2026)
Only accuracy changes: it moves from 6 (vendor run, off the board) to 8 (independent board result above Jev on Full; mixed on JevBench Intelligence, so not 9).
- Capability 7.0 → 7.4 = types 10 × 0.30 + calibration 5 × 0.25 + latency 6 × 0.25 + accuracy 8 × 0.20 = 3.00 + 1.25 + 1.50 + 1.60 = 7.35 → 7.4. Calibration stays 5 (board axis 77.8, no ECE, below Sage 1.4’s 83.7 which scored 6). Latency stays 6 (hosted p99 5.8 s outweighs the board’s 27 ms on-GPU p50).
- Maturity 6.6 and adoption 22 are unchanged.