Fastino explains how it trained GLiDE v1.2
Fastino has published the training recipe behind its GLiDE v1.2 decision model, showing how it moved chain-of-thought reasoning into the weights to cut latency more than 100-fold while still improving accuracy.
Fastino has published a technical write-up explaining how it trained GLiDE v1.2. As we reported earlier, the model topped decision-model leaderboards. The post, updated November 6, says v1.2 ranked #1 on the Hugging Face Decision Index v0.4.1 at its October 22 launch, with a score of 71.5. It also shares the top displayed score of 65 on the Artificial Analysis Decision Index.
According to Fastino, GLiDE v1.2 is post-trained from an open-weight 27B-class hybrid model that has 48 Gated DeltaNet linear-attention layers and 16 full-attention layers. It returns calibrated probabilities over typed options in a single forward pass and generates no tokens.
The central technique is distillation. The first GLiDE was a reasoning model that scored 64.81, but its mean latency was 42.6 seconds. Fastino then used the reasoning model's answers, along with other teachers, as soft probability targets for a student that never produces thought tokens. The first student scored 62.47, up from about 58.7 for the base model with reasoning off. GLiDE v1.1 later passed its teacher at 65.12 with more than 100× lower mean latency. On the same index edition, v1.2 scores 71.12. Learned pause tokens, implicit chain-of-thought internalisation and re-running middle layers did not beat this plain approach, the company says.
Fastino says its largest per-benchmark gains came from rebuilding training data to match how evaluators format their questions. On the ACOS aspect-sentiment benchmark, matching the evaluator's question structure and its 0.56% "yes" base rate raised the score from 27.2 to 61.6. Oversampling the weak benchmarks had raised it only to 30.1.
The post also describes routing teachers by skill, but the excerpt available to us ends before the details.