Side by side

The same six questions, the same answers as the two model pages, in two columns.

How to read this

Nothing here is written separately. Every answer below is the short answer from the Jev or Laya page, set beside its opposite number; those pages carry the rest. Where a column is short, that is the finding: one of the two has published nothing.

What kind of model is it

TypeSafe has published no answer to this one; Convai has.

Jev

Unknown.source

TypeSafe's launch post says they built “a new model architecture, parallel sampler for maximum efficiency, and training method we call Reinforcement Learning for Calibrated Decisions (RLCD)”. That is the whole of it. The words transformer, attention, encoder, decoder and parameters appear nowhere in the post or the docs.source

Laya

An encoder.config.json

A BERT-family transformer that reads the whole input in both directions at once and produces no text at all. config.json says ModernBertForMaskedLM: hidden size 1024, 28 layers alternating full and sliding attention, mean pooling.config.json

How it reads the state

Jev

About 64,000 tokens per request, of which 32,000 may be the state.source

How it represents them is not published.source

Laya

512 tokens per question on the English checkpoint, 1,024 on the multilingual one.model card

Separately, the decision head has its own budget for the options: head_max_len is 192 tokens on English, 256 on multilingual. That number matters, and question 6 explains why.model card

How it turns options into numbers

TypeSafe has published no answer to this one; Convai has.

Jev

Not published.source

TypeSafe describes a “parallel sampler” that produces every answer in one query rather than token by token, which tells you the shape of the computation and nothing about the mechanism. Up to 255 options per choice question are supported.source

Laya

Every option you send gets its own [MASK] token.model card

The model scores each of those markers, then softmaxes across them, and that distribution is the answer. This is the part worth reading twice.model card

How it was trained to be calibrated

TypeSafe has published no answer to this one; Convai has.

Jev

RLCD, Reinforcement Learning for Calibrated Decisions.source

The name is published. The method is not: no reward function, no data, no procedure. Laya's authors use the same name for their own training and do publish the details, so the Laya page can describe a concrete algorithm where this page can only report a name.

Laya

RLCD, Reinforcement Learning for Calibrated Decisions, with the method published.model card

The policy reports a probability distribution. Exploration adds zero-mean Gaussian noise to the logits. The reward is a strictly proper scoring rule (log and spherical, plus ranked probability score for ordinal questions). Updates are REINFORCE with a group-mean baseline, the same idea as GRPO.model card

What it costs to run

Jev

Access
A paid API, from TypeSafe directly or through Vercel AI Gateway. Signups were paused shortly after launch.source
Price
$0.042 per million input tokens. Output is free, because there is no generated text.source
Speed, stated
70 to 500 ms per requestsource
Speed, measured
p50 of about 305 ms through the Gateway, network includedmeasured here
Hardware
None of yours. It runs on TypeSafe's.

Laya

Price
Free. Apache 2.0. You pay for hardware, and nothing leaves your machine.model card
Install
pip install laya. Weights, about 2.3 GB, download from Hugging Face on first use.
Speed, stated
39.5 ms for one question on a T4 GPU, 32.8 ms on the multilingual checkpointmodel card
Speed, measured
p50 107 to 185 ms across the three games, on a laptop CPU with no GPUmeasured here
Three checkpointsmodel card
CheckpointBackboneParamsContextAimed at
layaModernBERT-large421M512English, guardrails, email triage
laya-multilingualmmBERT-base322M1,024100+ languages, about 2.2x faster
laya-typed-decisionsModernBERT-large421M1,024the four typed-decision workflows

What it cannot do

Jev

  • Run on your machine.No weights, no local option, and your state leaves your network on every call.
  • Be fine-tuned.Whatever it knows is what you get.our inference, from there being no published training path
  • Be inspected.You cannot check any claim about it against anything.
  • Be relied on for throughput.We hit 1,231 rate-limit retries (HTTP 429) and 14 outright failed decisions across our runs.measured here It can also return 529 when overloaded.source
  • Return an invalid option.Worth stating as a genuine strength: it cannot answer off the list. It can still be wrong.

Laya

  • Handle many options well.The options share that fixed head_max_len budget, so 77 options get roughly 3 tokens each and the labels blur together. On Banking77, Convai report Laya at 0.425 against Jev's 0.870.model card The limit is a consequence of the mechanism in question 3, not a mystery.
  • Match Jev on soft distributions.Higher argmax accuracy on typed decisions (0.766 against 0.727) but lower soft accuracy (0.471 against 0.580).model card
  • Arrive calibrated out of the box.Its 0.081 ECE comes after domain temperature fitting. Raw, the base checkpoint is worse than Jev.model card

What we measured

Three games, the same seeds and the same text to both models, everything recorded. The per-metric table is on the Scorecard, with the raw numbers on Results.

Jev

Three games, zero training, 10 episodes each.measured here

  • Highway:survived all 40 seconds, every episode. No crashes.measured here
  • Blackjack:matched basic strategy on 77% of decisions (95% CI 73 to 81).measured here
  • Snake:ate 1.8 food per episode, against 17.3 for a 15-line greedy script.measured here
  • Reading the situation:its answers moved substantially between opposite situations, which is the test for whether it is reading the state at all rather than answering from habit.measured here
  • Caution:it kept its lane on 94% of highway decisions and drove near the minimum allowed speed. In Snake it circled rather than dying.measured here
  • So: it does not crash, it does not do anything stupid, and it does not win. Safe and unambitious is a real decision-making style, and on a highway it is the right one.

Laya

Three games, zero training, 10 episodes each, on CPU.measured here

  • Blackjack:chose stick on all 200 decisions, scoring exactly like an always-stick baseline. Matched basic strategy 49.5% of the time, which is what you get by not deciding.measured here
  • Highway:64% of its moves were lane changes into a lane that did not exist. Average survival 18.1 seconds against Jev's 40.measured here
  • Confidence runs backwards:when it claimed 80 to 90%, it was right about 55% of the time; when it claimed 90 to 100%, about a third.measured here
  • The fairness note.This is a fair result for these games and an unfair one for the model. Laya was built for support tickets, moderation and routing, not for driving. On its own published benchmarks it beats Jev on accuracy (0.766 against 0.727 on typed decisions, 0.950 against 0.910 on AG News) and is three times better calibrated (ECE 0.081 against 0.246).model card, and Convai note that the Jev figures there are third-party published, not measured by them
  • What these games measure is not which model is better. It is how far each one travels from the domain it was trained on, and games are much further from Laya's home than from Jev's.
  • Its weights are open, so the obvious experiment is to fine-tune it on decisions like these and put it back in the arena. That is the next thing on the list.

Sources

Both models are days old and changing. Everything here was checked on 23 September 2026 against the sources listed above.