Six Decision Models in Three Weeks. Three You Can Own.

Last updated: 14 October 2026 — extended to cover six models.

The Short Answer

Six decision models shipped in three weeks, from TypeSafe's Jev on 15 September 2026 to OpenAI's Decisions API on 6 October. Three carry Apache 2.0 weights. The split from generation is an architecture now, and on ibl.ai you own all the code and the data — the router included.

Most enterprise agent stacks still send every task to one frontier model. Classify the ticket, check the permission, pick the knowledge base, judge the confidence, decide whether to escalate — then, finally, write something. Five of those six steps produce no prose at all.

What is a decision model, and why is it not just a smaller LLM?

A decision model returns a structured answer rather than text. There is no decoding step, because there is nothing to decode.

Given a passage and a set of typed questions, GLiNER2.5-Decide returns valid answers with probabilities, confidence scores and constraint-feasibility metadata. It can extract spans and relations and enforce rules across related outputs.

It cannot write you a paragraph, and it is not trying to.

That is the architectural point. Routing, triage, tool selection and guardrail checks are the frequent judgment calls inside an agent pipeline, and they have been running on generation models for the same reason everything else did: that was the only thing in the stack.

Which decision models shipped, and what licence does each carry?

Six, inside three weeks, with the same thesis and radically different distribution.

Jev publishes no air-gapped deployment option; the three Apache 2.0 entries do, because the weights are on your disk.

When one vendor argues the decision layer should be separate, that is positioning. When six ship inside three weeks and the largest incumbent follows, it is an architecture.

Which decision model is the 33 ms one, and is it called Vega?

The 33 ms model is Laya, from Convai Innovations, released 18 September 2026 under Apache 2.0. Vega is a real model, but a different one, and the two are being merged into a single claim that describes neither.

Laya is the one that matches the description doing the rounds — a state plus typed questions in, typed answers with calibrated probabilities out, one forward pass, no token generation.

It is an encoder, not a shrunken LLM: the English checkpoint is 421M parameters on ModernBERT-large, the multilingual one 322M on mmBERT.

And the 33 ms is the smaller one. On a Tesla T4, Convai reports 39.5 ms per question for the 421M English checkpoint and 32.8 ms for the 322M multilingual variant — so the rounded 33 ms belongs to the lighter model, not the flagship.

Where that figure comes from is worth a sentence, because it is a pattern you will meet again. Convai's own project page leads with "32.8 milliseconds on a single GPU" and never says which GPU; the Tesla T4 appears only on the model card.

A latency number without its hardware is not a measurement. The same 32.8 ms would be unremarkable on an H100 and impressive on a T4, and a buyer cannot tell which claim was made. Ask for the checkpoint and the chip, every time.

The model card is unusually candid about the rest, and this is the part that never travels with the latency figure. On zero-shot typed decisions Laya scores 0.362, against 0.318 for random guessing and 0.461 for always picking the majority class.

Read that again: out of the box it loses to a constant. The card says so itself — "Laya is a fast base to specialise, not a zero-shot decision engine" — and the fine-tuned laya-typed-decisions checkpoint reaches 0.766 on the same measure.

That is a genuinely useful model and an honestly documented one. It is not a drop-in judge, and a buyer who deploys it on the strength of a latency number will be measuring the wrong thing.

What is Vega, then, and who publishes it?

Decision-2.0-Vega-27B is the largest member of the Decision 2.0 family from the vLLM Semantic Router team, announced 2 October 2026 under Apache 2.0. It is a real model, and it is not the 33 ms one.

The family is a graduated range released together, spanning Kai at 0.6B through Eos at 0.8B, Sol at 2B, Nox at 4B, Lux at 9B and Vega at 27B — all answering several typed questions about one input in a single forward pass and returning a probability per option.

Vega itself is 29.37B parameters, an adapter on Qwen3.8-27B, with a 32,768-token context and a median 71.4 ms per single-question request on one GPU. It reports 74.0 on JevArena and 56.5 on the Jev Decision Index.

Note that the model card does not name that GPU either. The figure is still useful as an order of magnitude — roughly twice Laya's T4 latency for roughly seventy times the parameters — but it is not a number to put in a capacity plan.

The more interesting result is what the team did with the range. A cascade from Kai to Vega, published 10 October 2026, gained 16.45 percentage points of accuracy over Kai alone at an estimated 54% lower inference cost than Vega alone on JevBench.

That is the decision layer getting its own router. The cost argument that pushed classification off frontier models applies again one level down, and it only works if you hold every checkpoint — which an Apache 2.0 family lets you do and a metered API does not.

How fast is GLiNER2.5-Decide, and on what hardware?

Fastino's GLiNER2.5-Decide is fast — but the number depends entirely on the machine, and this is where the figure circulating online goes wrong.

Fastino publishes p50 end-to-end latency on short documents across a range of hardware: 38.3 ms on an NVIDIA V100, 43.4 ms on an L4, 43.6 ms on a T4, 47.3 ms on an A100, and 167.3 ms on a 48-vCPU Intel Xeon Platinum 8581C.

At 1,024 tokens the same measurements rise to 52.6 ms on an A100 and 75.6 ms on a V100.

Both halves of that table matter, and a summary that takes the fast number and the CPU claim together gets the model wrong. The sub-40 ms figures are GPU measurements; on the 48-vCPU Xeon it is 167.3 ms, roughly 4.4× slower than the V100.

That the model runs on a CPU at all is the genuinely interesting claim. It just does not run at GPU latency there.

Has anyone benchmarked these decision models against each other?

Not on neutral ground — and the numbers being passed around suggest otherwise, so it is worth stating clearly.

Fastino reports GLiNER2.5-Decide at 60.1% average across 17 datasets spanning classification, routing, triage and content understanding, against JevK5 at 57.5%, SemIf at 56.4%, GLiFormer at 49.0% and Laya at 46.6%.

JevK5 is not TypeSafe's Jev. It is an independent open-weight reproduction of the idea, built by a third party on Qwen3.5 and released under Apache 2.0, and its own repository states that it is not affiliated with TypeSafe AI or Jev. TypeSafe's model does not appear in these results at all.

Two further caveats, both from Fastino's own write-up: the suite is its own internal benchmark, not an independent one, and the baselines are its own selection.

Laya's 46.6% here is at least consistent with its own model card, which puts the base checkpoint below a majority-class baseline until it is fine-tuned. A 421M encoder measured out of the box is being measured before the work that makes it useful.

Vega's 74.0 on JevArena and 56.5 on the Jev Decision Index come from its own model card, which footnotes them as a run of the official 0.2.1 kit rather than the public-only Decision Index 0.3.1 figure Drex 1.5 is ranked on — so the two are not like-for-like.

The family is on the board, though: pull request #48 against the Decision Index repo records its 0.2.1 runs being folded into the board's v0.3.

So every figure in this category is a vendor scoring itself against alternatives it chose — useful as a sanity check that small models are competitive at this task, and not a basis for picking between them.

So what actually separates them?

The licence, and it is not close.

GLiNER2.5-Decide is Apache 2.0, and Fastino documents it running locally on CPUs and in air-gapped environments. Laya and Decision 2.0 are Apache 2.0 too. That makes the decision layer a component you hold rather than a service you call — and three weeks ago it was not.

The consequence is not philosophical. A router you rent can be repriced, deprecated or version-bumped underneath a workflow you have already validated.

It also sees every routing decision your organization makes, which for a regulated buyer is a record of internal operations leaving the building even when no customer data does.

Is there a leaderboard for decision models now, and does Drex 1.5 top it?

There is a board, a third model has arrived, and the rank-1 claim travelling with it was withdrawn by the vendor that made it.

Nace AI released Drex 1.5 on 28 September 2026 and opened its weights around 9 October. The widely-shared version of this story says it took #1 on the Decision Index 0.2.1 with 58.28, surpassing Jev 1.13.0 at 57.91. Four things are wrong with that sentence.

The board moved on. The live board is Decision Index 0.3.1, and Nace AI's own pull request against its results repo, merged 10 October, is titled "Report Decision Index 0.3.1 (public only) for Drex v1.5" and replaces the 0.2.1 figures.

The vendor narrowed the rank. On 0.3.1 the public-only scores are Drex v1.5 58.08, Jev 57.96 and Bespoke Nimble 9B v3 57.19.

Nace AI's own PR describes Drex as "level with Jev," and the overall rank-1 claim is gone. What a later commit in the same repo asserts instead is that "Drex v1.5 is the #1 model under 10B parameters."

The margin is inside the published tie band. The Decision Index kit shares a rank between models less than 0.9 points apart, on the grounds that a gap that small is what a fresh sample of questions would produce 95% of the time. The widely-shared pair differs by 0.37.

It is the wrong number. The board ranks on a Full score that is 20% public benchmarks, 50% private tests of the same skills and 30% private tasks from new domains. Drex has a self-run public-only figure, and the kit's own README notes that "a public-only advantage counts for little."

Worth naming what the Decision Index is, too. It was built to reproduce TypeSafe AI's Jev benchmark suite, so the scoreboard for this category is currently named after one competitor's model and run by neither vendor nor an independent body.

The defensible claim is narrower and still interesting: Drex 1.5 is the strongest sub-10B model on the public subset of Decision Index 0.3.1, self-reported.

What licence do Drex 1.5's open weights actually carry?

A restricted RAIL licence, which is why "open-sourced" is the wrong word for it, and the distinction is the whole argument of this post.

The weights are genuinely downloadable, in bf16 and as a Q8_0 GGUF, from the model card. The licence is the NACE AI OPEN RAIL-M LICENSE (MODIFIED), not an OSI-approved one.

It bars commercial use by any entity with more than $1,000,000 in gross revenue in the prior twelve months, or that has raised more than $1,000,000 in total equity or debt funding. It also bars use in "any product or service that competes with" the licensor's.

It also bars using the model "to make fully automated decisions that produce legal or similarly significant effects on individuals… without meaningful human review" — which rules out unsupervised high-stakes deployment, though routing and triage with a human in the loop remain permitted.

Set that beside GLiNER2.5-Decide under Apache 2.0 and the licence gradient in this category is now visible across three models: one Apache 2.0, one hosted-only and closed, one open-weight but commercially fenced.

Which is the point. Downloadable is not the same as ownable, and the test is whether the licence survives contact with your revenue, your sector, and your use case.

How many projects actually use Jev, and what is the 2,170 figure?

2,170 is a count of public GitHub repositories as of 22 September 2026, not applications, and it accumulated in one week rather than three.

The number comes from an academic survey, Jev in the Wild by Guoming Ling, Muen Xue and Zijian Ye, which collected 2,170 publicly available Jev projects. Of those, 1,865 were created in the week following the 15 September release and 305 were existing repositories that added Jev.

The distinction between a repository and an application matters for a buyer. A week-old repo is a developer trying something; the paper's own framing is "projects," and it uses that interchangeably with repositories.

What the survey does show is what people reach for a decision model to do. Attribute judgment appears in 77% of projects, scoring or ranking in 52%, action selection in 31%, and 69.7% use it for two or more purposes at once.

That last figure is the architecturally interesting one.

Most adopters are not replacing one classifier — they are moving a cluster of small judgments out of the generation path together, which is precisely the shape that makes the decision layer worth owning rather than renting per call.

Worth noting what the paper does not contain: any percentage split of enterprise AI between classification and generation. The claim that "80% of enterprise AI is classification, not generation" circulates without attribution, and we are not going to put a number on it.

The qualitative version is defensible and we have argued it before — most of what an agent does all day produces no prose — and that argument stands without an invented statistic.

What changed when OpenAI shipped a Decisions API?

The category stopped being a startup story, and the ownership question got sharper rather than softer.

OpenAI opened its Decisions API to all developers in public beta on 6 October 2026, three weeks after Jev defined the category.

It returns typed answers to developer-defined questions — predicates with a probability, choices from a fixed set with confidence, and scores against ordered levels — over text and image input.

It runs on GPT-6 Luna, and OpenAI puts it at about 10× faster than calling the same model through the Responses API, priced at $0.10 per million input tokens with output and cache reads unbilled.

Compare that to Jev's $0.042 per million input tokens and the pricing story is now a two-vendor race to the bottom on a metered API. That is good news for anyone buying decisions by the call.

It is also the clearest possible demonstration of why the licence column is the one to read.

Two of the six are hosted endpoints from well-capitalized vendors, and both can be repriced, rate-limited, deprecated or quietly re-pointed at a different base model. Image input is a real capability advantage; it is not ownership.

Three weeks took this from one product to six, and from zero downloadable options to three under Apache 2.0. The buyer's position improved — but only for buyers who took the weights.

What does this change about what an enterprise should build?

Stop budgeting as though one model does everything, and keep the decision layer replaceable.

An agent workflow makes many structured judgments per generated response. Pricing that whole shape at frontier rates is how AI budgets end up dominated by work that never needed a large model.

Splitting the layers is the fix, and it is now a choice among six shipped implementations rather than a thing to build yourself.

The part worth protecting is optionality, and the last three weeks are the argument for it.

This post began with two models nine days apart; by the time OpenAI's endpoint opened on 6 October there were six, a leaderboard had been revised, a rank-1 claim had been narrowed by the vendor that made it, and the most-quoted latency figure in the category turned out to describe a checkpoint nobody meant.

Nothing in that list is a reason to avoid the layer. All of it is a reason not to hard-wire a single supplier into it.

A platform that lets you swap the decision layer without rewriting the agents around it is worth more than any current benchmark leader — because in this category the current benchmark leader has a shelf life measured in days.

A seventh entry landed a week after this post, and it is the one that put a brand name on the idea.

We took it apart in Constrained Decoding Models Are Not a New Architecture, including the vendor's own benchmark table — which puts its paid hosted extraction model 0.2 percentage points ahead of GLiNER2.5.

Why does ownership decide this one?

Because the router is where your operating logic lives, and a rented router is somebody else's copy of it.

On ibl.ai you own all the code and the data. The platform is model-agnostic, so the decision layer is a component you choose and can change.

That includes an open-weight model running entirely inside your own network, which is exactly what an Apache 2.0 model that runs air-gapped makes possible.

Pricing is usage-based with no per-seat pricing, and you can deploy anywhere: your cloud, your VPC, on-premise, or fully air-gapped.

More than 1.6M users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.

ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.

Sources: model, licence, hardware latency table and the internal benchmark from Fastino's launch post; Jev's pricing and announcement date from TypeSafe's launch post; JevK5's independence from its own repository; Laya's parameter counts, Tesla T4 latencies, zero-shot scores and the "fast base to specialise" caveat from its model card, with the unqualified 32.8 ms claim on Convai's project page; Vega's size, licence, context, latency and scores from its model card and the cascade result from the vLLM Semantic Router blog; Drex's licence from its model card; the 2,170 repository count and the purpose breakdown from Jev in the Wild.

Related: Generation Is Commoditized. Judgment Is the New Frontier — Jev, the first model in this category, in depth, and who owns the rubric · Most University AI Work Is Judgment, Not Generation — where the budget actually goes when you split the layers.