Fastino trains AI to make decisions without generating text

Fastino’s GLiDE v1.2 scores 71.5 on the Decision Index and yields probabilistic decisions in a single pass without text generation.

When a business needs to identify a suspicious transaction, choose an agent's next tool, or route a support ticket, it does not necessarily need an AI model capable of writing several paragraphs. It needs to select from defined options and provide a reliable estimate of how likely its decision is to be correct. Fastino Labs is pursuing this approach with GLiDE v1.2, a specialized model that accepts text, JSON, and images and returns probability distributions directly, without generating text. The model scores 71.5 on Hugging Face's Decision Index, with a reported median latency of 93 milliseconds in the published evaluation. Its weights are available on Hugging Face, alongside a hosted API for organizations that do not want to operate their own inference infrastructure. Returning Probabilities Instead of Written Answers

General-purpose language models can classify documents, select tools, and decide which action an application should take. They must usually generate a response, and sometimes an extended reasoning trace, even when the required output is simply a category or Boolean value. That generation introduces computation and latency for every decision.

GLiDE, short for Generalized Lightweight Decision Engine, operates differently. The model receives a context followed by one or more questions with predefined possible answers. It supports multiple-choice decisions with up to 255 options, yes-or-no questions, and ordinal scoring. Its inputs can combine written text, structured JSON, and images.

Instead of producing prose, the model returns probabilities for the available choices. A customer service application might use it to decide whether a request concerns billing, technical support, or security. A software agent could select an appropriate tool or determine whether an operation should require human approval.

The application receives a structured decision directly, without having to interpret generated text. It can then apply a confidence threshold, escalating uncertain cases or routing them to another system. Moving Reasoning From Generated Tokens Into Model Weights

Earlier versions of GLiDE relied on adaptive reasoning. When the model encountered a difficult question, it generated additional reasoning tokens before making a decision. The approach improved certain results but could introduce considerable latency. Fastino reports an average response time of 42.6 seconds for an intermediate reasoning-based version.

With GLiDE v1.2, the researchers sought to preserve some of those reasoning benefits without generating the reasoning itself. They trained a student model on probabilistic answers produced by models capable of explicit reasoning, allowing the student to make decisions in a single computational pass.

The process uses probability distillation, transferring not only a preferred answer but also the teacher models' uncertainty across possible options. The student learns to distinguish straightforward questions from ambiguous ones through its internal representations rather than by writing out its reasoning.

According to Fastino's research report, the initial single-pass model recovered part of the earlier reasoning gains. After several training iterations, the student surpassed its teacher on the team's selected evaluations. GLiDE v1.1 reached 65.12 on one edition of the Decision Index, while v1.2 increased that score to 71.12 on the same edition.

The team also experimented with learned pause tokens, implicit reasoning techniques, and repeated execution of intermediate layers. None outperformed its conventional distillation approach. The released design therefore uses one forward pass without intermediate text generation. A 27B-Class Model Specialized for Decisions

GLiDE v1.2 is post-trained from an open-weight model in the 27-billion-parameter class, using a hybrid architecture with 48 Gated DeltaNet linear-attention layers and 16 full-attention layers.

Development involved several stages: introducing reasoning capabilities, distilling them into direct decisions, adapting training data to real evaluation protocols, and improving probability calibration.

One finding was that some errors arose not because the model lacked the necessary ability, but because its training examples did not match the way questions were presented during evaluation.

On ACOS, a benchmark for aspect-based sentiment analysis, the evaluation asks numerous binary questions about potential category-sentiment combinations in customer reviews. Only 0.56% of those questions have a positive answer. Fastino's original training examples had a much higher positive rate, causing GLiDE to identify far too many aspects per review.

By rebuilding the examples to match the evaluation format and natural distribution of answers, the researchers report improving ACOS from 27.2 to 61.6. The result illustrates how the structure and statistical distribution of training questions can strongly influence performance. Combining Teachers With Different Strengths

GLiDE v1.2 does not rely on a single teacher during distillation. Fastino uses previously released decision models, an earlier GLiDE checkpoint, and a frontier language model serving as an annotator.

Their contributions are weighted by task family. A teacher performing well at tool selection may be less useful for language analysis or knowledge-intensive questions. The team therefore combines their predictions using a weighted product of experts, with weights fitted on held-out examples.

When the combined teacher prediction disagrees with a known correct answer, the reference label takes priority. The resulting probabilities become training targets for the student.

Fastino reports a 1.45-point accuracy improvement over the best individual teacher in its out-of-fold evaluation. Some teachers receive almost no weight on particular tasks, while an older checkpoint can remain valuable when dealing with unfamiliar domains.

These training-time combinations do not require every teacher to run when