Fastino GLiDE v1.2 Claims #1 on the Decision Index at 71.5: What Is Verified

On October 22, 2026, Fastino Labs announced GLiDE v1.2 and called it "the world's best decision model," claiming first place on the Hugging Face Decision Index with an overall score of 71.5. The announcement on X also says v1.2 has one of the lowest calibration errors on the index, adds image input, ships weights on Hugging Face, and costs $0.04 per million input tokens through Fastino's API.

The 71.5 figure is Fastino's own claim, and the weights come with a licence that is not a standard open-source one. This post checks what we could verify against the Hugging Face model repository and Fastino's pages, and what remains a vendor claim.

Claimed vs verified at a glance

What is GLiDE, and how is it different from a chatbot?

GLiDE is a decision model: you give it a state and a set of typed questions, and it returns probabilities. Per the README, there are three question types: choice (one of up to 255 named options), noul (yes or no, returning the probability of yes) and score (ordered levels). Nothing is generated. Each question takes one forward pass, and the response reports output_tokens: 0.

Here is the README's example. A support ticket says a customer was charged twice. Asked what the customer wants, the model returned refund with probabilities of 0.9751 for refund and 0.0229 for other, a "should this be escalated today" answer of 0.8267, and a priority score at level 2 with 0.9239 probability. The point is a typed, machine-readable answer that code can act on, with a calibrated confidence, rather than free text to parse.

Fastino built the model on a Qwen-family base: the repository tags list qwen3_5 and image-text-to-text, and the weights total about 26.1 billion bf16 parameters. That is a large model next to Fastino's earlier GLiNER2.5-Decide, a 340M encoder under Apache 2.0 that runs on CPUs. GLiDE trades that lightness for accuracy on harder decisions with long context, up to 40,000 tokens per question by default.

An earlier version of GLiDE was described by Fastino as the first "thinking decision model": it produces a fast probability distribution, then spends extra reasoning on uncertain choices. See Fastino's introduction to GLiDE, which reported 64.81 on Decision Index 0.2.1 against 57.91 for Jev, on Fastino's own run with the official scorer and with the public leaderboard paused for submissions.

How credible is the 71.5 score?

Treat it as a vendor claim until a third party reproduces it. Three points matter.

It is Fastino's number. The X post, the website chart and the earlier blog post all come from Fastino. The website chart we could read lists Perplexity Decider v1.2 at 68.0, vLLM-SR d3 at 67.2 and Jev 1.14 at 60.2, with GLiDE at 71.5. We did not find the Hugging Face index page confirming it.

"The Decision Index" is not one fixed thing. Different publishers use versions and variants. In our coverage of Base Labs, mammoth-1 claimed 71.24 on the Jev Decision Index 0.4 and first place among 208 models, as we described in that claimed-vs-verified post. A 71.5 on one index version and a 71.24 on another are not directly comparable, and two labs claiming first place within days of each other tells you the indexes differ or the submissions differ, not who is ahead.

Calibration is claimed, not shown. Fastino says GLiDE's confidence more closely reflects how often its answers are correct. That is the right property for a decision model, since downstream code uses confidence to route or escalate. But the post does not give a calibration number, and we could not find the underlying method. If calibration matters to you, measure expected calibration error on your own labelled data.

Can you actually run it?

The weights are downloadable, but the practical bar is high.

The server speaks Fastino's System One format on POST /v1/systemone and batches questions across requests. For background on that format, see our explainer on System One models.

The hosted route is simpler: Fastino's API, at $0.04 per million input tokens according to the announcement. Since the model emits no tokens, the bill is input only. For comparison with other priced options, our Liquid AI D1 Ultra write-up covers a rival at $0.042 per million input tokens, and the Perplexity Decider API post covers another.

What does the licence actually allow?

This is the biggest gap between the announcement and the repository. The X post says Fastino is "excited to release the model weights on Hugging Face." The repository, however, tags the licence as other, includes a LICENSE file, and its README says "Internal use only." The README also says the download needs a token with read access to the fastino organization.

We did not read the full LICENSE text, so we cannot say precisely what is allowed. The safe reading is that "weights you can download" is not the same as "open source", and that commercial redistribution, fine-tuning for resale, or embedding in a product may be restricted. That contrasts with models such as GLiNER2.5-Decide under Apache 2.0 and Base Labs' MIT-licensed mammoth-1. Read the licence before you plan around the weights.

How does GLiDE v1.2 compare with other decision models?

Three podium steps with a green stone on the tallest, illustrating competing first-place claims for GLiDE v1.2 on decision model indexes

The table uses only figures from sources we cite, and mixes index versions, so read it as a map, not a ranking.

Notice that Perplexity Decider v1.2 gets 68.0 in one vendor's chart and 69.42 in another's. That is a small illustration of why self-run, differently versioned indexes cannot settle a ranking. For a wider tour of the field, see our top decision models ranking and the Clairic, Jev and OpenAI comparison.

What are people saying?

Replies under Fastino's post were light-hearted. One user asked whether it could eventually be installed locally, and another joked that a 71.5 score is solid but they would still trust a friend who is wrong most of the time for life choices. A third asked whether Fastino would add the image bounding-box feature of another Hugging Face model, vLLM-SR d3-turbo. These are small samples from a post with about 13,000 views and not evidence of adoption.

Which model for which job?

What should you do before adopting it?

1. Run your own labelled test set through the hosted API with the $0.04 per million token price in mind. A few hundred examples beat any leaderboard.
2. Measure calibration, not just accuracy: bucket confidences and compare them with actual hit rates.
3. Read the LICENSE file in the repository before you download or deploy the weights.
4. Check whether you need 80 GB of GPU memory, or whether the hosted API is enough.
5. Wait for an independent entry on the Hugging Face index before citing 71.5 as a ranking.

We will update this post if a third party reproduces the score or Fastino clarifies the licence.

Related reading

• What are decision models?
• GLiNER2.5-Decide: a 340M open-weight decision model
• Base Labs mammoth-1 and elephant-1: claimed vs verified
• Top 10 decision models ranked
• Liquid AI D1 Ultra: claimed vs verified
• Perplexity Decider and the Decisions API
• What is a System One model?

Details reflect Fastino's announcement and the fastino/GLiDE repository as of October 23, 2026. Scores, prices and licence terms may change.