Fastino releases five-model GLiDE decision family
Fastino has released open-weight GLiDE decision models from 0.4B to 27B parameters, and all five sit on the Decision Index size–quality frontier, giving developers one interface across a wide range of costs.
Fastino has released GLiDE as a family of five open-weight decision models at 0.4B, 0.8B, 2B, 9B and 27B parameters. All five share one interface, the System One schema, so developers can change model size without changing their questions or how they parse the answers. As we reported earlier, GLiDE v1.2 had already reached #1 on decision-model leaderboards. This release extends the line across five sizes.
Fastino cites the official Decision Index 0.4.1 snapshot from November 6. On that board, no model with the same or fewer served parameters scores higher than any of the five GLiDE models. Full scores range from 33.84 for the 0.4B to 73.37 for the 27B. The 27B shares the tied #1 rank group with REAX 1.3, which scores higher at 73.98 but serves 57.36B parameters against GLiDE's 26.09B. The 9B scores 62.48, compared with 60.26 for Baseten's elephant-1.
GLiDE scores every option in a single forward pass and returns probabilities instead of generated text. Median latency on a single NVIDIA RTX PRO 6000 ranges from 7 ms for the 0.4B, which needs under 1 GB of memory, to 112 ms for the 27B, which needs 52 GB.
The 27B and 9B models use recurrent depth: a band of middle layers runs twice with shared weights. This adds computation without adding parameters or output tokens. At 27B, Fastino says it raised the expected Full score from 72.1 to 73.3 and increased median latency from 83 ms to 112 ms. The gain at 9B was smaller. The available source text cuts off partway through the 2B results.