Fastino releases GLiDE decision model
Fastino has released GLiDE, a lightweight model that returns structured, probability-scored choices from a fixed set of options and reasons further only when it is unsure, aimed at agent tool selection and action control.
Fastino announced the Generalized Lightweight Decision Engine (GLiDE), which it calls the first "thinking" decision model. GLiDE is available through the Fastino API.
GLiDE takes application state and a predefined set of actions. It returns one selected action, a confidence margin and a probability for every option. The model first produces a fast probability distribution. If the leading option is uncertain, it spends more compute on reasoning and folds that reasoning into the final probabilities. Fastino says this lets easy decisions stay fast while harder ones get handled more carefully, such as choosing among hundreds of tools, tracing code or applying policies that depend on several facts. The context window is 40k tokens, and thinking tokens count against it. Requests that are too large are rejected rather than truncated.
On the Decision Index 0.2.1, a benchmark of 38 tasks, Fastino reports a chance-adjusted score of 64.81, compared with 57.91 published for Jev. It says GLiDE leads in all five evaluation areas and on 31 of the 38 benchmarks. Reported gaps include:
- Knowledge and Reasoning: 62.9 vs 51.4
- Tools and Automation: 83.5 vs 75.1
- CLadder accuracy: 88.7% vs 72.6%
- CRUXEval accuracy: 92.6% vs 73.0%
Fastino ran these numbers itself using the official scorer. GLiDE is not on the public leaderboard, which is currently paused for 0.2.1 submissions.
The API follows the System One model schema, so applications can act on typed outputs directly instead of parsing generated text. Example uses include gating destructive agent actions, triaging production incidents and checking contracts against policy. Fastino did not disclose pricing, model size or latency figures.