Respan releases Span-02, an open-weight decision model that answers by a deadline

What Respan released on 15 October

Respan, the brand of Keywords AI, Inc., which sells tracing, evaluation and gateway tools for AI agents, announced Span-02 on 15 October 2026 under the tagline anytime reasoning for decisions. The company describes it as a model that sees images and screens, answers choices, yes-or-no questions, scores and numbers, and ships with open weights.

The weights went up on Respan's Hugging Face page alongside the source code of the server that runs them. According to the model card, Span-02 is a fine-tune of Google's Gemma 4 26B-A4B, a mixture-of-experts model with about 4 billion active parameters, quantised to a 16.4GB file. Respan also offers it as a hosted model through its own API.

What a decision model answers

Most language models are asked to write an answer that software then has to parse. A decision model skips that step. You send a state, which can be plain text or any JSON value plus up to four images, and a set of typed questions. For each question the model returns the answer together with its full probability distribution.

Span-02 supports four question types. A choice picks one option from a list of up to 679. A yes-or-no question, which the API calls noul, returns the probability of yes. A score rates something on an ordered scale of two to ten levels. A number returns the number itself with a confidence, so developers do not have to list candidate values as options. One request can mix all four, and the questions share the state, which the engine processes once.

The request format follows TypeSafe's System One API, so existing clients work by changing the URL; Respan describes its hosted endpoint as Jev-compatible. Jev is the decision model from TypeSafe, which we covered when the company announced an $870 million Series A. The number type, images and the reasoning settings are Span-02 extensions that those clients can ignore.

How the deadline works

Reasoning models usually let developers choose an effort level and then wait. Span-02 takes a deadline. Each request can set a reasoning mode, auto, on or off, and a budget in milliseconds from 0 to 600,000, where 0 means no limit. The model thinks until the deadline and then answers with its best decision so far and how sure it is.

The model card explains the mechanics. Each answer is first read directly from the model's probabilities for the option labels in a single pass. In auto mode, if the top answer's probability is below 0.8, the model writes a short thought and the answer is read again, and the final distribution is the average of the before and after readings. Thoughts are decoded with speculative decoding using Gemma's public draft model, at about 480 thought tokens per second on an RTX PRO 6000.

The thought itself is never returned unless a trace is requested on a self-hosted server, and Respan's documentation says thinking tokens are not billed on the hosted API. A per-question cap on thought length applies whatever the budget says.

What Respan's numbers show

Respan's main benchmark is Decision Index 0.3, a public suite it describes as 37 benchmarks scored on a chance-corrected scale where 0 is guessing and 100 is perfect. On Respan's own local rebuild of the suite, Span-02 scores 61.2 with thinking off, 63.0 with a one-second budget, 63.6 with three seconds and 64.1 with no limit. Respan's chart sets that against 58.0 for Jev 1.13.0, 61.7 for Cloudflare's Clef, 62.2 for Perplexity's Decider and 49.0 for Liquid AI's d1-3B. Respan notes that the longer-budget figures add the change measured on a 5,000-question sample to the thinking-off score, and that latencies were measured differently for each model.

The model card gives a text index score of 63.62 and a vision score of 70.69 on the public index, with every request answered. Respan also shows two demonstrations. In one, the model is asked whether each of 4,050 points on a 4-degree grid is on land, from latitude and longitude alone; accuracy rises from 71.9 percent with thinking off to 84.7 percent with one second, against 79.6 percent for Jev. In the other, Span-02 picks links in a Wikipedia speed-run from Alexander the Great to the University of Milan in 20.6 seconds, against 38.5 seconds for GPT-6 Luna, though Respan notes Luna's goal included a hint.

Running Span-02 on your own GPU

Self-hosting has strict requirements. The weights run only with the included engine, a Rust and CUDA server, and cannot be loaded with Hugging Face transformers or vLLM. The kernels use instructions specific to Nvidia's Blackwell generation, so the card lists only the RTX PRO 6000 Blackwell with 96GB and the RTX 5090 with 32GB, plus a driver with CUDA 12.9 support. Memory use is about 28GB with default settings.

The quantisation mixes formats: the mixture-of-experts layers use 4-bit NVFP4 weights and activations, attention and dense layers use 8-bit FP8, and the vision tower and router stay in bf16. Respan provides a Dockerfile and a launch script that downloads the draft model and starts the server, which it says is ready in about 20 seconds.

On Respan's measurements on one RTX PRO 6000, a stratified sample of public text questions took a median of 71 milliseconds and a mean of 308 milliseconds per request, rising to a median of 616 milliseconds on a set where 78 percent of questions think. With 64 requests in flight it reports 167 requests per second without thinking on the RTX PRO 6000 and 127 on the RTX 5090, and says its engine is 1.3 to 7 times faster than vLLM on the same RTX 5090 without thinking.

Using the hosted API

For developers without a Blackwell card, Span-02 is available at Respan's decisions endpoint with a Respan API key. The docs say it runs on Respan credits only and is priced per input and output token; the model page lists $0.01 per million input tokens after gateway discounts and a 131,000-token context. Calls without credits return an error, the model has its own per-organisation rate limit, and streaming is not supported.

Respan's examples are typical decision-model jobs: routing a support ticket to billing or tech support, checking whether a customer is asking for a refund, rating frustration on a scale, or reading how many days late an order is. Because Span-02 sees screenshots, Respan also pitches it for computer-use agents that need to pick the next click quickly.

Where Span-02 fits among decision models

Span-02 lands in a crowded field of models that return typed decisions rather than prose. Cloudflare's Clef-omni reads audio and video, and Liquid AI has published open-weight d1 models for edge devices. Several of them are compared on the same Decision Index, which makes the claims easier to line up, though each vendor still runs and reports its own numbers.

What Respan adds is the deadline. Production agents often have a hard latency budget, and a model that can trade a few hundred milliseconds of thought for a better answer, while still returning something useful if time runs out, fits that constraint better than a fixed effort setting. The calibrated probabilities matter too: an agent can act on a confident answer and escalate an uncertain one.

What to keep in mind

All the benchmark, latency and throughput figures are Respan's own, from its rebuild of the public suite and its own hardware, and the demonstrations were chosen by Respan. AI Lookout has not run the model. Developers should check whether the confidence scores are well calibrated on their own tasks before relying on them to decide when to escalate.

Open weights here come with practical limits. The model is tied to Respan's engine and to recent Nvidia hardware, so it cannot be dropped into a standard inference stack, and number questions think only in the always-on mode. The weights are a derivative of Gemma 4 and distributed under its Apache 2.0 terms, while the engine is Apache 2.0 with credits to vLLM, SGLang, FlashInfer and other projects in the repository's notice file.