Today we are releasing GLiDE v1.2, the newest version of our Generalized Lightweight Decision Engine. GLiDE v1.2 ranks first on the Decision Index v0.4.1 with a score of 71.5, leading the next-ranked model by 1.7 points. You can run inference through the Fastino API. Weights are available on Hugging Face.

Hugging Face Decision Index v0.4.1. October 22, 2026.

Most of what businesses ask language models to do isn’t writing. It’s deciding. Is this ticket urgent? Which tool should the agent call, or should it ask a follow-up question first? Does this claim match the filing? Should this transaction be flagged? The answer is one of a known set of options, and what you really want is a probability you can put a threshold on.

GLiDE is a decision model built for exactly that. You give it a context (text, JSON or images) and one or more typed questions: multiple choice over up to 255 options, yes/no, or an ordinal score. It returns a calibrated probability for every option, from one forward pass, without generating a single token.

We removed thinking tokens to improve latency, while significantly boosting accuracy after understanding how decision models are used at scale today. This blog covers how we got here and what we learned, including what didn’t work.

Try GLiDE - the #1 model on Decision Index

Get started

Fifteen days since

On September 15, Jev launched: a closed, hosted decision API and the model that turned “decision model” into a product category. Nobody outside its makers knew how it worked. A week later, Hugging Face’s Decision Index went live, and Jev was at the top. The index is a deliberately broad benchmark: 38 benchmarks and 155,390 requests, spanning PhD-level science (GPQA), AGI-level questions from Humanity’s Last Exam, tool calling, retrieval, contracts, clinical trials, chess and even humour.

Twelve days after the launch of Jev, we launched our first version of GLiDE, which had higher accuracy than Jev on a full run of the official harness: 64.14 to 57.91. On September 30, we released an updated version of GLiDE that scored 64.81.

Then benchmarks for decision models got more complex as use cases expanded. On October 6, the Decision Index introduced hidden test sets and removed reasoning from the main leaderboard. Under this updated evaluation, GLiDE was one of the first two models to score above Jev on the Decision Index.

A week after that, Artificial Analysis launched its own Decision Index. It combines four tests: niche knowledge questions with an “I don’t know” option, Humanity’s Last Exam, tool-calling decisions and safety classification. GLiDE placed second on the list of the eleven decision models and had the lowest calibration error of any of them.

Timeline from Jev’s launch to GLiDE v1.2

How we trained GLiDE v1.2

GLiDE v1.2 is post-trained from an open weight 27B-class hybrid language model, with 48 Gated DeltaNet linear-attention layers and 16 full-attention layers. Getting from the base model to v1.2 took four stages. The first three built the foundation; the fourth is where v1.2 was made.

1. Teach it to reason, then distill the reasoning into the weights

Everyone knows chain-of-thought helps models with hard questions. Fewer people know you can take the thinking back out.

The first version of GLiDE was a reasoning model: when its one-pass confidence was low, it thought before answering. We bootstrapped that reasoning in two steps. First, we fine-tuned on reasoning traces whose answers we could verify. Then we used rejection sampling to distill the thinking model into itself. It scored 64.81. But hard questions triggered long thoughts, and its mean latency was 42.6 seconds.

So we distilled it again, this time into a model that answers in one forward pass. The reasoning model’s answers after thinking, alongside other teachers, became soft probability targets for a student that never produces a thought token. Whatever reasoning helps, the student has to do it in its hidden states: in latent space, not in text.

It works better than you’d expect. With reasoning switched off, the original model scored about 58.7. Our first single-pass student scored 62.47, recovering about 60% of what thinking had bought, with zero thinking tokens. A few rounds later, the student overtook its teacher. GLiDE v1.1 scored 65.12 at more than 100× lower mean latency, and v1.2 scores 71.12 on the same edition.

| The thinking doesn’t disappear. It moves from tokens into weights.

We also tried giving the single-pass model explicit room to “think” internally: learned pause tokens, implicit chain-of-thought internalisation, and re-running middle layers. None of them beat the plain approach. What worked was soft targets from teachers that do think.

2. Copy the exam, not the answers

Our single largest per-benchmark gains came from fixing how we asked questions. On six benchmarks where we trailed the best open model by 8–37 points, we rebuilt held-out items exactly the way the evaluator builds them, and they reproduced our public deficit almost exactly. The model had the skill. It had never seen the protocol.

Take ACOS, an aspect-sentiment benchmark. The evaluator asks about every category–sentiment pair for every review: up to hundreds of yes/no questions per review, 0.56% of them “yes” across the benchmark. Our training rows asked about a handful of pairs, and 39% of their answers were “yes”. So the model predicted 11.8 aspects per review when the truth was 1.27. Rebuilding the rows to match the evaluator’s decomposition and its natural base rate moved ACOS from 27.2 to 61.6. Earlier, simply oversampling the weak benchmarks had moved it only to 30.1.

3. Route teachers by skill

Soft targets carry a teacher’s uncertainty, which is where a student’s calibration comes from. But no single teacher is best at everything. So for every task family, we combine several teachers: three publicly released decision models, our own earlier checkpoint, and a frontier LLM annotator. We combine them as a weighted product of experts, with the weights fitted out of fold on held-out rows. If the pooled answer agrees with the gold label, the pool is the target. If it disagrees, gold wins.

Out of fold, the routed targets beat the best single teacher by 1.45 accuracy points. The router learns sensible things: one teacher is switched off entirely on tool decisions, and our earlier checkpoint gets the largest weight on new-domain rows.

4. Teach it to say “I don’t know”, from its own knowledge

A model that always answers will sometimes answer wrongly with confidence. The Artificial Analysis index rewards abstaining instead, and so do real users. The trick is that only the student knows what the student knows, so abstention targets can’t come from a teacher.

We author niche closed-book questions from openly licensed documents and score the starting model on them. We then pick the confidence threshold that best trades off accuracy against hallucination and build targets from it: confident and correct means answer; correct but unsure means mostly answer; wrong means “I don’t know”. “I don’t know” is just another option. There’s no special head and no inference-time rule.

Artificial Analysis Decision Index components, launch checkpoint versus GLiDE v1.2

Non-hallucination went from 16 to about 78. One lesson came the hard way: when we added new knowledge later but kept abstention targets computed before the model learned it, accuracy rose by 6 points and non-hallucination fell by 5. Abstention targets have to reflect what the model knows now.

How we kept ourselves honest

Public benchmarks are easy to overfit, so we designed the evaluation process to reward improvements that generalize by:

• Calibrating our internal evaluations against published results.

• Separating model selection from final confirmation.

• Testing performance on an independently created domain.

• Monitoring a broad range of real-world capabilities for regressions.

The final checks could reject a model, but they were never used to choose or refine one.

Results

Decision Index

GLiDE v1.2 is #1 on Decision Index 0.4 with a Full score of 71.5, 1.7 points ahead of Jeffan-31b-1710 and 11.3 ahead of Jev. The Full score combines private tests of the same skills (60%) with private tasks from new domains (40%).

The lead holds on both parts. GLiDE v1.2 scores 68.3 on same-skill tests and 76.3 on new domains, the highest of the top eight on each. Across the five private areas, it leads on knowledge (57.3), retrieval (73.1) and tools (84.9). Two models beat it in one area each: mammoth-1 on language and Jeffan-31b-1710 on arts. It also has the lowest calibration error of the top seven (2.0) and a 93 ms median latency on the board.

On the public suite, which everyone can run, GLiDE v1.2 scores 71.5 in our complete run of the official harness. That’s 6.4 points up from v1.1 and 14 points ahead of Jev 1.13, with a lead on 32 of the 37 benchmarks. It trails on GPQA Diamond, MMLU-Pro, Humanity’s Last Exam and API-Bank. Those reward recalling rare facts and long multi-step reasoning, which is exactly what thinking tokens buy. That’s the frontier we’re working on next.

Calibration

Because GLiDE outputs probabilities, they should mean something. Across 4,482 held-out decisions, GLiDE v1.2’s expected calibration error is 1.9% from a single forward pass, with no temperature tuning.

On the same items, three open decision models achieve 76.9% to 78.5% accuracy. Large open reasoning models, read out after thousands of tokens of thinking, had 12 to 16% calibration error before rescaling.

Speed

One forward pass per question gives an 82 ms median and 245 ms at the 80th percentile on a single RTX PRO 6000, in 52 GB of GPU memory. A full request with images runs at a 93 ms median. The model never generates a token, so there’s no output-token cost and no sampling. The same request on the same hardware gets the same answer every time.

What didn’t work

We tried far more than we kept. A few highlights from the 25 dead ends in our technical report:

• Patching a finished model with short fine-tunes: the patching itself cost hidden-set points, whatever the content. Retraining from the parent was better.

• Architecture changes that cost latency: bidirectional recurrent layers (+15% latency) and depth recurrence (+25%) bought nothing.

• More of the same benchmark data: bulk training splits raised public scores while lowering our hidden-set estimate.

• FP4 quantization: quantizing only the MLP layers made it 29% faster but cost 0.35 points; quantizing every layer cost more. We ship bf16.

Try GLiDE v1.2

Get started with GLiDE today.

GLiDE v1.2 is available to download on Hugging Face under the Fastino Community License, and for inference today on the Fastino API.

For information on pricing and getting started, please refer to our docs.

Check out the Hugging Face Decision Index.