Span-02: a decision model whose thinking time is adjustable to the millisecond

Respan launches Span-02, an open-weight multimodal decision model capable of scaling its reasoning to a precise deadline and providing probabilistic responses.

An immediate decision in a few milliseconds, or more time to examine a complex situation. With Span-02, Respan offers a decision model whose reasoning duration can be set directly by the developer. Unlike systems that require choosing a predefined effort level, Span-02 accepts a budget expressed in milliseconds and returns the best available decision at its expiration. Multimodal and distributed with open weights, it supports binary responses, classifications, scores, and numerical values, for an advertised price of $0.01 per million input tokens. Setting a time limit rather than a reasoning level

Models with reasoning capabilities generally offer several effort levels, which determine the amount of compute devoted to a response.

Span-02 introduces another method, called anytime reasoning.

The developer can set a maximum reflection time in milliseconds. The system uses this time to progressively improve its decision, then returns the best result obtained when the time limit expires.

An application can thus request an almost immediate response or allow for a longer reflection when a situation requires more analysis.

Three modes are provided.

The off mode disables additional reasoning and prioritizes speed.

The on mode enables this capability with a budget configurable through the `budgetms` parameter.

The auto mode lets the system determine whether it should continue its reasoning based on the uncertainty associated with the decision.

The Respan documentation indicates that the budget can be set between 0 and 600,000 milliseconds. A value of zero means that no time deadline is imposed, although internal limits still apply.

This setup allows adapting processing time to the constraints of an application rather than systematically deploying the same reasoning effort. Structured decisions rather than written responses

Span-02 belongs to a category of models specialized in situation analysis and response selection.

Instead of producing primarily paragraphs, it returns structured results accompanied by confidence indications.

The system supports four question formats.

Choice allows selecting from several pre-defined options, with a probability distribution for the different options.

Noul corresponds to a binary decision, whose response is expressed as a probability.

Score places an observation on a scale, for example to measure the urgency of a request or the risk level associated with a situation.

Number directly produces a numerical value, accompanied by a confidence level.

This last feature sets Span-02 apart from several decision models limited to classifications and probabilities.

It becomes possible to ask how many items appear in an image or to extract an amount without having to define all possible responses in advance.

All four types can be combined in a single request, based on a shared context. A score that improves with reflection time

Respan evaluated Span-02 on Decision Index 0.3, a public benchmark suite comprising 37 benchmarks designed to measure the capabilities of models specialized in decision-making.

With reasoning disabled, Span-02 achieves a score of 61.2, compared to 58.0 for Jev 1.13.0 in the comparison presented.

Enabling reasoning progressively improves the results.

At 270 milliseconds, the score reaches 61.5. It rises to 62.5 with 500 milliseconds, then to 63.0 after one second of reflection.

Respan also announces a score of 64.1 when the model can continue its reasoning without a predefined deadline.

These measurements describe a trade-off between speed and decision quality: longer processing can improve the result, without guaranteeing an identical progression for every question.

The scores come from a local reproduction of the public suite carried out by Respan. For the longest reasoning durations, the company also uses an estimation of gains measured on a sample of 5,000 questions.

This is therefore not a fully independent evaluation of all advertised configurations. Latency comparisons to be interpreted with caution

Respan positions Span-02 among decision models capable of responding quickly, including when reasoning is enabled.

The lab indicates that with this feature disabled, its model outperforms Jev while being approximately four times cheaper.

It also claims a latency roughly three times lower than that of certain models displaying a comparable performance level.

However, the comparison conditions are not perfectly uniform.

Span-02's response times are measured on Respan's infrastructure, while some competitor data comes from API requests and includes network delays.

The published graph notably distinguishes between server-side measurements and round-trip times from other providers.

Differences in hardware, load, and environment can influence the results.

The speed observed by a developer will therefore depend on the chosen reasoning mode, the infrastructure used, and the complexity of the transmitted information. Drawing a world map from coordinates

To illustrate the effect of reasoning time, Respan submitted a geographical experiment to the model.

The system receives only latitude and longitude coordinates. It must determine whether each point is located on land or in water, without access to a map.

The experiment covers 4,050 points distributed over a four-degree global grid.

Without additional reasoning, Span-02 achieves an accuracy of 71.9%. With 100 milliseconds, it reaches 72.5%.

At 300 milliseconds, the result progresses to 84.5%, then to 84.7% after one second.

The graphical representation resulting from these responses gradually reveals the outlines of the main continents.

Jev 1.13.0 achieves 79.6% on the same question, according to Respan's measurements.

Researchers use Natural Earth geographical data as a reference to determine the presence of land or water surfaces.

This experiment illustrates the potential improvement in responses given extra time, but does not constitute a general measure of the model's spatial reasoning capabilities. Images and interfaces among the analyzed data

Span-0