On 28 September a 144M-parameter model called julia-1 appeared that did something unusual: instead of generating an answer, it scored a fixed set of options and returned the best one. One forward pass, no prose.
Six weeks later that is a product category with entries from Perplexity, Microsoft, Cloudflare, Liquid AI, Aleph Alpha and several independents, present on Hugging Face, Ollama and OpenRouter, and absorbing 138 billion tokens a day on the last of those.
This is what has actually shipped, and what it does and does not add up to.
The list
| Model | Params | Date | Notable |
|---|---|---|---|
| julia-1 | 144M | 28 Sep | The first — scored options, no generation |
| Cloudflare Clef | 27B | ~2 Oct | Typed decisions; the release that named the category |
| autotrust JEV-27B-VL | 27B | ~3 Oct | Vision-language; benchmarked against Clef |
| Aleph Alpha Kolibri-1 | MoE | ~4 Oct | German-language reasoning |
| autotrust GEV-26B-Decide | 26B | ~5 Oct | Adaptive thinking budget on Gemma |
| convaiinnovations laya | 421M | ~6 Oct | Non-autoregressive; published ECE of 0.081 |
| Liquid AI d1-3B | 3B | 5 Oct | Multimodal, LFM2-VL based, ar/zh/en |
| Phocinae-Largha | 144M | 8 Oct | First encoder; zero output tokens |
| Microsoft Decision-1 | — | ~10 Oct | A hyperscaler enters |
| Perplexity pplx-decider-v1.1-27b | 27B | 6 Oct | On OpenRouter, 138B tokens/day |
| strands-decider | — | ~12 Oct | Entered Ollama’s popular list |
Eleven in sixteen days, if you count from Clef.
What they genuinely agree on
Strip the marketing and there is a real shared architecture, and it is a good idea.
The answer space is enumerable. Yes/no, pick-one-of-N, score 2–10. Because the set of possible answers is finite and known in advance, the model does not have to write anything — it scores the options and you read the scores. Phocinae’s usage field reads {"input_tokens": 146, "output_tokens": 0}, which is literally true rather than a figure of speech.
One forward pass. No autoregressive loop, so latency is a single inference rather than a token-by-token generation. Phocinae reports 21 ms p50 on an RTX 5090; laya is in the same territory.
Confidence comes out with the answer. This is what makes the category more than a cheap classifier. If the model returns a calibrated probability alongside its choice, you can build a gate: handle the confident cases locally, escalate the rest to a real LLM. Phocinae measures that gate at τ=0.6 cutting LLM calls 55% while raising kept-subset accuracy from 0.906 to 0.9936.
The target workload is agent interiors. Command approvals, tool selection, step checks, output screening — the repetitive internal decisions of an agent session, each of which is currently a 500–4,000 token API call returning roughly one token of information.
That last point is why this appeared now rather than two years ago. The category exists because agents turned out to make enormous numbers of tiny decisions, and paying frontier-model prices for each one is absurd.
Where the numbers stop being comparable
Here is the part the category has not been honest about collectively, even where individual cards have been.
Specialists are being compared to generalists. Phocinae reports 0.906 on typed-decisions against laya 0.766, meraGPT 0.768 and JEV 0.727 — and its own card states that Phocinae was fitted on that dataset’s train split while the others were measured zero-shot. That is a fine-tuning gap partly masquerading as an architecture gap. The card says so; the comparison still circulates without the caveat.
“Calibrated” is doing different work in different places. laya published an ECE of 0.081 on raw output. Phocinae published 0.2519 shipped and 0.0168 with its bundled calibration column — a fifteenfold correction, applied at inference. Both are legitimately described as calibrated models. They are not describing the same property, and a τ=0.6 gate on an ECE-0.25 model routes the wrong cases.
Most of them publish no calibration figure at all. Of the eleven above, two have measured ECE. The confidence-gating argument is the whole economic case for the category, and it is unverifiable for most entries.
Option-order invariance is measured by exactly one of them. Phocinae reports flip robustness of 0.0200–0.0283. Nobody else reports it. A pick-one model whose answer changes when you reorder the options is exhibiting a positional artefact rather than making a decision, the bias is well documented in LLM multiple-choice evaluation, and the usual mitigation — averaging over permutations — is unavailable inside a 21 ms production gate.
The claim nobody has defended
Every model in this category is sold on the same promise: a small model can make these decisions as well as a large one, because the decisions are simple.
Nothing published tests the interesting version of that. The benchmarks are datasets of typed decisions, and they measure accuracy on decisions that were already formatted as typed decisions by someone who knew the answer. What is missing is the harder question: when a decision looks simple and is not, does the small model know?
That is exactly what the confidence gate is supposed to handle, and exactly what an ECE on an in-distribution test set does not measure. A model well calibrated on cases resembling its training distribution can be confidently wrong on the case that actually matters — the command that looks routine and is destructive, the tool that is semantically adjacent and wrong. Phocinae’s own card, to its credit, publishes a known defect of exactly that kind: Router.route_tool() confuses semantically similar tools.
It also fails JevBench public-231 at 0.5455, below the 58.4% gate, and says so. Across all eleven, that remains the only published result that undercuts its own case.
What to do with this
If you are building something that makes repetitive structured decisions, the architecture is sound and the economics are real. A local 150M encoder answering a gating question in 21 ms, deterministically, with nothing leaving the machine, is a genuinely better component than a round trip to an API — and for creative work on client material, the determinism and the privacy are worth as much as the latency.
The things to demand before adopting one:
- A measured ECE on the shipped configuration, not on a corrected one
- An option-order flip figure
- Whether the headline accuracy is zero-shot or fitted on the benchmark it is quoting
- Your own eval, built from your own decisions, because the published benchmarks measure a distribution that is not yours
Our guide to building that eval covers the arithmetic. Fifty cases from your own history will tell you more than any number in the table above.