Is GLiDE state of the art?Pending benchmarks.Announced Nov 5, 2026; no independent results yet.
Fastino's open-weight GLiDE decision-model family: five sizes (0.4B, 0.8B, 2B, 9B, 27B) that share the System One schema (choice, noul yes/no, score) and return a probability for every option in one forward pass with no generated reasoning tokens. GLiDE 27B and GLiDE 9B use recurrent depth (a band of middle layers loops twice with shared weights). Weights on Hugging Face under the Fastino Community License; GLiDE 27B is also hosted on the Fastino API (api.fastino.ai System One, model id fastino/GLiDE).
The details so far
- Lab
- Fastino
- Announced
- Nov 5, 2026
- API
- Live · GLiDE 27B on Fastino API POST https://api.fastino.ai/v1/systemone with model fastino/GLiDE; API key required
- Open weights
- Published
- Parameters
- 0.4B / 0.8B / 2B / 9B / 27B (family)
- Role
- Non-autoregressive decision models: one forward pass per question over a state (text, JSON, optional images) with a probability for every supplied option
- Sizes
- GLiDE 0.4B, 0.8B, 2B, 9B and 27B; all share the System One schema so questions and parsing stay the same across sizes
- Question types
- choice (up to 255 options), noul (yes/no), score (ordered levels); multiple questions per request
- Recurrent depth
- GLiDE 27B and GLiDE 9B run a band of middle layers twice with shared weights; 2B and below ship without it
- Context
- Up to 40,000 tokens per question prompt by default on the 27B card (state, question, options and image tokens)
- Modalities
- Text and JSON; images supported (lab: Recurrent Looped GLiDE / 27B path accepts images)
- Checkpoints
- fastino/GLiDE-0.4B, GLiDE-0.8B, GLiDE-2B, GLiDE-9B, GLiDE-27B on Hugging Face (collection fastino/glide); BF16
- License
- Fastino Community License (HF license tag: other); third-party components retain their licenses
- Predecessor
- Separate from Fastino GLiDE 1.2 (glide-v1-2) and the earlier GLiDE no-thinking 28B board row (decision-27b-ckpt)
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: on Decision Index 0.4.1, expected Full scores (80% interval) are 73.3 (71.4–75.2) for 27B, 61.8 (57.4–65.5) for 9B, 44.5 (40.6–47.7) for 2B, 36.2 (32.3–39.4) for 0.8B and 30.3 (25.2–33.9) for 0.4B.
Lab claim: public-suite scores measured with the official harness are 72.5 (27B), 68.8 (9B), 60.6 (2B), 54.3 (0.8B) and 51.1 (0.4B); median latency on one RTX PRO 6000 is 112 / 39 / 18 / 19 / 7 ms respectively.
Lab claim: GLiDE v1.2 had the lowest calibration error of any decision model on the Artificial Analysis Decision Index board (prior generation, cited in the family post).
Sources
LABThe GLiDE Family blog (Fastino)fastino.ai/blog/glide-decision-model-family · Nov 5, 2026REPOfastino/glide collection on Hugging Facehuggingface.co/collections/fastino/glideREPOfastino/GLiDE-27B model card (Hugging Face)huggingface.co/fastino/GLiDE-27BLABFastino on X: GLiDE familyx.com/fastinoAI/status/2118405436919648618 · Nov 5, 2026Tracking continues until an independent board lists it.