GLiDE
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity4.9/10
Self-serve hosted API (api.fastino.ai/v1/systemone) with a full reference, pricing and a TypeSafe migration guide. Closed weights, no GA statement, no SLA. Fastino's docs say it is not a drop-in replacement for Jev (Score semantics differ), so SDK reuse is partial.
- Capability6/10
Noul / Choice / Score, up to 255 options and 40k tokens. Adaptive thinking adds reasoning tokens on hard items, and no latency figure is given. Confidence is a top-1 minus top-2 margin, with no calibration metric. Full Decision Index run (64.81) is self-run with the official scorer, not on the board.
- Adoption13/100
Day-one launch with solid reach on X (~270 and ~140 likes on the two main posts, ~60k views combined). Fastino already has a hosted decision model in the catalog. No integrations or production cases reported for GLiDE yet.
Vendor claims
- GLiDE never trained on a row of Decision Index data; Fastino run puts pplx-decider at 56.39 and Clef-flash at 57.1 (Fastino employee post, 2 Oct 2026) — UnverifiedClaim[Vendor claim — not independently verified]
- Decision Index 0.2.1 score of 64.81 vs Jev's published 57.91, using the official scorer in Fastino's own complete run (not on the public leaderboard)[Vendor claim — not independently verified]
- Leads Jev in all five areas and on 31 of 38 benchmarks; Knowledge and Reasoning 62.9 vs 51.4; Tools and Automation 83.5 vs 75.1[Vendor claim — not independently verified]
- CLadder 88.7% vs 72.6% and CRUXEval 92.6% vs 73.0%[Vendor claim — not independently verified]
- The first thinking decision model[Vendor claim — not independently verified]
Read this first
- 64.81 is Fastino’s own run with the official scorer, although the launch post from @george_onx says “Ranks #1 on the Decision Index”. GLiDE is not on the official Decision Index, where submissions to 0.2.1 are paused. Cloudflare claimed “leader” for Clef (61.21) on the same day, also self-run.
- “Thinking” means the model generates hidden reasoning tokens on uncertain items. They appear as
output_tokensand are billed at $0, but they add latency, and no latency numbers are published.- Not a drop-in for Jev. Fastino’s docs say so:
scorereturns an integer index plusexpected_level, and confidence is the margin between the top two options.- “First thinking decision model” is a marketing claim. TypeLLM has offered a priced “thinking” mode for decisions since 29/09.
GLiDE (“Generalized Lightweight Decision Engine”) is a decision model from Fastino, introduced in a blog post dated 30 September 2026 by Sahibzada Allahyar and Mary Newhauser, and launched on X on 1 October (15:27 BRT, @fastinoAI and @george_onx). It is hosted only: no weights, no parameter count, no base model disclosed.
GLiDE first produces a fast probability distribution over the options. If the leading answer is uncertain, it spends extra reasoning and folds it into the final probabilities. Fastino pitches this for agent controllers choosing among hundreds of tools, multi-step reasoning (arithmetic, dates, code tracing) and verifying other models’ output.
Specs
| Attribute | Value |
|---|---|
| Maker | Fastino |
| Weights | Closed (hosted only); size and base undisclosed |
| Endpoint | POST https://api.fastino.ai/v1/systemone, model fastino/glide |
| Auth | X-API-Key or Authorization: Bearer |
| Question types | Noul, Choice (up to 255 options), Score |
| Context | 40k tokens per question; oversized requests are rejected, not truncated |
| Price | $0.30 / M input tokens; thinking tokens reported as output and priced at $0 |
| Fine-tuning | Not available for GLiDE yet (docs: no fine-tuned deployments) |
| Launch | (blog date) |
Evidence (UnverifiedClaim)
| Area | GLiDE | Jev 1.13 (published) |
|---|---|---|
| Decision Index 0.2.1 | 64.81 | 57.91 |
| Knowledge and Reasoning | 62.9 | 51.4 |
| Tools and Automation | 83.5 | 75.1 |
| CLadder | 88.7 | 72.6 |
| CRUXEval | 92.6 | 73.0 |
Fastino says GLiDE leads Jev in all five areas and on 31 of 38 benchmarks. Mary Newhauser’s thread adds +6.9 on HLE.
A follow-up chart from Fastino (posted 2 Oct by @singularity_bly, whose profile says he works at Fastino; the post itself does not say so) adds Clef 61.2 (self-reported), Clef-flash 57.1 and pplx-decider 56.39 run by Fastino. The y-axis starts at 56, which makes the gaps look larger than they are. One useful check: the public board has pplx-decider at 56.40, so Fastino’s harness matches the board within 0.01 on that model. That makes the self-run 64.81 more credible, but it is still not on the board. The claim that GLiDE “never trained on a row of DI” is unverified. No score change.
Fit / anti-fit
Fit: hard, reasoning-heavy decisions where a few hundred extra milliseconds are acceptable (agent guardrails, tool selection over long menus, verification); teams fine with a hosted API.
Anti-fit: latency-critical paths with no tolerance for variable response time; self-hosting or data-residency needs; code that relies on Jev’s exact Score semantics.
Limits
- No latency figures, no calibration metric.
- The launch post says “inference and training”, but the docs list no fine-tuned GLiDE deployments yet.
- Fastino’s pricing page and model page disagree on GLiNER-2.5-Decide’s price ($0.03 vs $0.15 per M), so check prices in the console.
Why it is in the catalog
GLiDE is a model Fastino trained for decisions and serves through a documented typed API that returns probabilities. The thinking step is internal; the caller still gets structured answers, not text. It sits next to Fastino’s existing GLiNER-2.5-Decide.