Fastino open-sources GLiDE decision models: 27B scores 72.49 on Decision Index at 112ms
RobertTLange · x · 2026-11-09
Fastino released GLiDE, a family of five open-weight decision models (0.4B–27B) built for fast, calibrated structured decisions: given app state plus a set of candidate actions, one forward pass returns the chosen action, confidence, and per-option probabilities — no thinking-out-loud tokens.
Key technical details:
- Built from Qwen3.8-27B: 3:1 Gated DeltaNet/attention hybrid made bidirectional; LM head + MTP dropped (−1.7B), replaced by a 255×5120 readout over option letter codes
- "Thinking" via recurrent depth: layers 24–47 run twice (½x + ½G(x)), 64 → 88 layer passes with zero new params
- 40K tokens/question, up to 255 options, supports text/JSON/images
- Decision Index 0.3.1: 72.49 for 27B, 51.10 for the 0.4B encoder; 112ms median latency on a single RTX PRO 6000
Free under $10M annual revenue via the Fastino Community License.
More from Models
- Tiny fine-tunes are trending on Hugging Face: 0.8B Qwen model runs bash translation offline on CPU — victormustar · 2026-11-09
- Commenter recants: NVIDIA may not save Western open source after all — teortaxesTex · 2026-11-09
- World model funding cools sharply: just 5 deals in October, VCs shortlist only 4-5 teams — teortaxesTex · 2026-11-09
- User says Terra Max's batch browser use is the main reason for their ChatGPT subscription — amplifiedamp · 2026-11-09
- DeepSeek V4.1 Flash 1017 hits 1,000 tok/s on Vercel while a home 4x 3090 rig does 65 — AccBalanced · 2026-11-09
- Grok usage limits reset within hours after Musk-led engineering review — XFreeze · 2026-11-09