Fastino opens GLiDE v1.2 weights, and it takes #1 on the Decision Index

Source · Model card: GLiDE

Fastino released GLiDE v1.2 on 22 October 2026 (@fastinoAI, 13:36 BRT), this time with weights on Hugging Face (fastino/GLiDE) alongside its API. Until now GLiDE was hosted only.

What shipped

• Model: post-trained from Qwen3.8-27B, a hybrid of linear-attention and full-attention layers. choice (up to 255 options), noul and score over text, JSON and images, 40,000 tokens per question, one forward pass, bf16 weights of 52 GB.
• No more thinking tokens. The first GLiDE reasoned on hard items. Fastino’s write-up says that reasoning was distilled into a single-pass student that later beat its teacher.
• License: Fastino Community License. Free to use, but a company with more than US$10M in annual revenue needs a commercial agreement for production use, and commercial users must display “Built with Fastino GLiDE”. Not an OSI license.
• API price: $0.04 per 1M input tokens, down from $0.30.

What the boards say

• Decision Index 0.4.1: #1 with 71.52, 1.77 ahead of Jeffan-31b-1710 and 3.08 ahead of mammoth-1. Jev 1.13 is at 59.22.
• JevBench API board (v1.6.40, 22 Oct): Capability 84.7 (#2 of the APIs), Intelligence 76.4 against Jev 1.13’s 63.6. Three items Fastino had already seen were left out of the run.

Fastino also reports 1.9% calibration error and an 82 ms median on one RTX PRO 6000 (UnverifiedClaim). The card still says “Internal use only” although the repository is public and ungated.

The catalog page has a draft re-score; see GLiDE.