Skip to content
Predictive Systems
PSI Daily

System 1 Decision Models Share One Recipe. OpenJev, Kev, and Laya Are Proof the Recipe Travels.

InterpretabilityAllan C. Tan, MS

TypeSafe’s Jev made mute decisions famous. OpenJev, Kev, Laya, and peers are now shipping the same System 1 contract in the open. The breakthrough is a shared architecture builders can host, measure, and improve.

Three matte decision tokens and a small open gear on dark slate under amber side light, suggesting a shared open decision mechanism

The contract common to all of them

Strip away the brand names and the systems rhyme. TypeSafe’s Jev, APUS OpenJev, Jared Palmer’s Kev, ConvAI’s Laya, and the growing set of open recreations all pursue the same System 1 contract.

You pass a state (text, ticket, email, or JSON) plus typed questions. The model returns structured answers with probabilities. It does not write an essay. The three primitives keep showing up: Choice (pick one option and a distribution), Score (place the state on an ordinal rubric), and Noul or yes/no (a calibrated P(true)). Schema violations are hard by design because the output space is numbers, not free text.

Under the hood, the shared algorithmic move is also clear. Skip token-by-token decode. Prefill or encode the state. Score a short candidate set in one forward pass. Emit probabilities software can branch on. Many teams then train or fine-tune against proper scoring rules (TypeSafe and Laya both use the RLCD label for calibrated decisions) so honesty in confidence is rewarded, not just top-1 accuracy. Kahneman’s System 1 framing is the marketing glue: fast reflex for routing, triage, guardrails, and agent tool choice; leave System 2 prose to generative models when you truly need it.

That common recipe is what makes the open wave interesting. Once the interface is stable, labs can compete on encoders versus decoders, latency kernels, multilingual routing, and eval harnesses without inventing a new product category each week.

OpenJev, Kev, and the decoder recreation lane

APUS AI Lab’s APUS-OpenJev-v1 (4B, 9B, 35B-A3B on Qwen 3.5, MIT tooling, Apache 2.0 weights) and Palmer’s Kev (0.8B, 4B, 9B) rebuild Jev-shaped local APIs from public behavior. Zhang Xu’s 36Kr interview stresses a functional recreation, not leaked TypeSafe weights. Vendor cards cite roughly 26 to 79 ms local decide paths on high-end NVIDIA GPUs, with about 9B as a practical quality floor on their ladders. Community harnesses such as Benchmark Heaven’s JevBench (pilot 19 Sep 2026) and a cluster of unrelated repos also branded OpenJev show the name is already a category nickname.

Laya and the encoder lane

Laya takes the same primitives down a different spine. ConvAI Innovations (Nandakishor Mukkunnoth) ships Apache 2.0 bidirectional checkpoints (ModernBERT-large about 421M, mmBERT-base about 322M), a script-aware multilingual router across 100+ languages, and claims about 33 ms p50 on a single GPU. Founder posts point to earlier arXiv work (2503.23303, 2510.01237) on schema decisions with RL. On published boards Laya looks strong on calibration and mid-size choice sets, weaker when option counts explode (Banking77-class heads starve under a shared token budget). That limit is itself an algorithm lead: shortlist, then decide.

What to build on next

Three open leads fall out of the shared architecture. First, decide kernels and batching so local Choice, Score, and Noul stay in the tens of milliseconds. Second, encoder versus decoder backbones, and how small a calibrated head can get. Third, public evals that report calibration and subgroups, not only accuracy, plus hierarchical choice for fat label spaces. Host an open path (OpenJev, Kev, or Laya) on your own tickets. Keep a hosted Jev lane as a reference if you need one. The mute core is no longer a single closed API. It is a portable contract the open ecosystem can own.

Sources

  • TypeSafe, Introducing System One Models and Jev: https://typesafe.ai/blog/introducing-system-one-models-and-jev
  • Simon Willison, 21 Sep 2026: https://simonwillison.net/2026/Sep/21/jev/
  • 36Kr / Zhang Xu (APUS OpenJev), 22 Sep 2026: https://eu.36kr.com/en/p/3993153099463681
  • APUS-OpenJev-v1: https://huggingface.co/apus-ailab/APUS-OpenJev-v1
  • Kev-9B: https://huggingface.co/jaredpalmer/kev-9b
  • Laya site: https://laya.convaiinnovations.com/
  • Laya on Hugging Face: https://huggingface.co/convaiinnovations/laya
  • JevBench v1.0: https://benchmarkheaven.com/jev-models/v1
  • arXiv 2503.23303, 2510.01237 (ConvAI prior framing)

Source: Laya