Intern-Decision-4B vs Laya: Two Models That Refuse to Write an Answer, Ten Times Apart in Size

There is a row in the Intern-Decision-4B model card that reads "Laya — 57.77". Anyone who has read Laya's own documentation will recognise the number's meaning: convaiinnovations/laya is a 421-million-parameter non-autoregressive decision model, and 57.77 is the average it earns when it is used zero-shot on somebody else's benchmark. Its own card says the same thing more bluntly — the base checkpoints sit below the majority-class baseline on the typed-decisions benchmark, at 0.362 against 0.461. Meanwhile internlm/Intern-Decision-4B is a 4.54-billion-parameter model built for exactly the same job: take a state and a set of typed questions, run one forward pass, return calibrated probabilities, never generate a token. Same architectural bet, same output shape, same Apache-2.0 license, ten times the parameters — and one of them is a base to fine-tune while the other claims to be a finished zero-shot engine.

They agree on the premise, which is the rare part

Both cards argue the same case, and it is worth stating because most of the industry argues the opposite. If your problem is choosing among a known set of answers, generating prose is the wrong operation: you pay for tokens you throw away, you need a parser, and you get an explanation you did not ask for instead of a probability you can threshold. Laya's card puts it as "it never generates text, so there is nothing to parse and nothing to hallucinate." Intern-Decision-4B's card says its API "performs structured candidate scoring. It does not call generate() or sample free-form text."

The mechanisms differ in an instructive way. Laya feeds typed questions to a classifier head and returns typed answers with calibrated probabilities in a single forward pass, with a routing layer that detects script and language in under half a millisecond before the pass. Intern-Decision-4B maps each question's options onto single-token symbols, renders an assistant JSON skeleton with one placeholder per field, reads the logits immediately before each placeholder, and softmaxes over only that field's allowed symbols. Laya's is a classification problem; InternLM's is a next-token problem it refuses to let become a generation problem.

The size gap, and what it buys

Asking two models to do the same task at 421M and 4.54B parameters produces the result you would predict, then a surprise.

The predicted part is capability on hard questions. Intern-Decision-4B averages 90.02 across seven evaluation sets with a Brier score of 0.347 and an expected calibration error of 0.065 in InternLM's own measurement, and it runs 44.16 ms mean per query on a single RTX 4090. Laya's base English checkpoint is 0.362 accuracy on the 2,000-decision typed-decisions benchmark, which its own card flags as below the 0.461 majority-class line and barely above the 0.318 random baseline. When the same checkpoint is fine-tuned on that benchmark's own training split the number jumps to 0.766. Laya's card draws the conclusion itself: "Laya is a fast base to specialise, not a zero-shot decision engine."

The surprise is that the smaller model wins two dimensions outright. Speed: Laya answers 103 to 332 questions per second batched on a single T4, and its card puts it roughly six to eight times faster than TypeSafe Jev on the independently measured 236–276 ms p50 figure it cites. Ten times the parameters does not buy ten times the throughput in the other direction — Intern-Decision-4B's 44.16 ms is per query on a 4090 with no batching story published. And language: Laya reports 45 of 51 benchmarked languages usable at more than three times random, with a routed 0.451 on MASSIVE intent in thirteen non-English languages against 0.306 for its English-only checkpoint. Intern-Decision-4B makes no multilingual claim at all.

Scoreboard

• Parameters — 4.54B BF16 for Intern-Decision-4B, text tower plus vision tower plus projector; 421M for Laya, a single compact encoder-class stack.

• Input ceiling — 8,192 tokens for both, and both refuse rather than truncate; note that Laya's 8,192 applies to the multilingual checkpoint, whose shipped default is 1,024.

• Multiplicity — Intern-Decision-4B takes 1 to 16 questions and up to 62 options each, and 62 is not arbitrary: it is exactly the number of single-token symbols its inference contract can address. Laya takes multiple typed questions too, but its card documents a sharp collapse past roughly 50 options, where a 77-label question gets only three or four tokens of budget per label and accuracy falls to 0.425.

• Images — up to eight for Intern-Decision-4B, counted against the 8,192-token budget; no multimodal path for Laya.

• Calibration — Intern-Decision-4B ships a fitted temperature of 1.99241824 selected by NLL minimisation over 1,728 cases, and reports ECE 0.065 on its own suite; Laya ships over-confident, and its card says refitting one temperature per question type and option count moves mean ECE from 0.466 to 0.081.

• Zero-shot posture — a finished checkpoint claiming a 90.02 average against a documented base that scores below the majority-class line.

• Latency — 44.16 ms mean, 44.03 ms median on one RTX 4090 against 103 to 332 questions per second batched on one T4.

• Languages — no published claim against 45 of 51 languages usable when routed.

• License — Apache-2.0 with the upstream Qwen license preserved against Apache-2.0 explicitly tagged for commercial use.

Two cards, two very different ideas of disclosure

This is where the comparison stops being a spec sheet and becomes a judgement call, because the two vendors document their models in opposite styles.

Laya's card publishes its own failures in detail. It reports that the English checkpoint scores 0.000 accuracy on Khmer at 0.952 confidence — confident while wrong, which makes confidence gating useless there. It reports that action.act_probability carries no usable signal, reading 1.0 for almost every input with an AUROC of 0.30 on 396 labelled decisions, and tells you to gate on confidence instead, which reaches 0.77 on the same items. It names ordinal score questions as "the weakest primitive", citing 0.372 on SST-5. A card that tells you which of its own outputs to ignore is doing something most vendors do not attempt.

Intern-Decision-4B's card publishes a clean table and no failure cases. Its caveats are real and load-bearing — the calibration section is unusually explicit that the "before" column in its pilot is not what any user would see, and that test-suite labels were not used to select the temperature — but the evaluation section is seven wins and no admissions. It also carries rows for five competing systems that InternLM ran on InternLM's harness, including the Laya row that opens this article. Those are not those projects' own numbers, and reading them as such would be a mistake.

So the sourcing line for this matchup is asymmetric in a way that favours the smaller model: Laya's evidence includes defects it documented itself, and Intern-Decision-4B's evidence has never met a test set its authors did not choose.

Which shape of problem each one fits

Given a short, structured state in English, a closed set of answers and a need for a trustworthy probability, Intern-Decision-4B is the more capable object on paper and the more expensive one in practice — four safetensors shards, PyTorch 2.9.1 and Transformers 5.14.1 under Python 3.12, a resident engine, and a wrapper you write yourself if you want an HTTP endpoint. Given a high-volume routing or guardrail decision across dozens of languages where throughput is the binding constraint, Laya is the better shape: pip install laya, a 421M model, and a card that has already told you to fine-tune before trusting the probabilities.

The one thing both cards agree on is that neither model should be trusted zero-shot without checking calibration on your own data — Laya because its base checkpoints are near chance on unfamiliar decision types, Intern-Decision-4B because its 0.065 ECE comes from a suite its own authors assembled.

What a router does and does not change here

Neither of these models is in the OrcaRouter catalogue, and it is worth saying plainly rather than implying otherwise: our model pages return a 404 for both Intern-Decision-4B and Laya, and neither is a route you can call today. What that means for the two projects is not the same. Laya's Apache-2.0 license with a commercial-use tag means the obstacle is purely packaging — it is a local Python package with a small footprint, which is the kind of thing that could plausibly be served. Intern-Decision-4B's obstacle is also packaging, but its 4.54B BF16 weights and 8,192-token refusal ceiling make it a component rather than a service.

Where OrcaRouter does help is on the far side of the decision. If your structured-decision problem turns out to need a reasoning step after all, the general models that can do it are on one key across 200-plus models, with provider list price passed through at 0% markup and automatic failover between providers — so the model you pair with a scorer is one endpoint away, not a second contract.

The short version

These two projects reached the same conclusion about how decisions should be made and then differed about nearly everything else. Laya bets that 421M parameters, tight language routing and ruthless honesty about its own defects make a fast base you specialise. Intern-Decision-4B bets that ten times the parameters and a carefully engineered single-token scoring contract make a finished zero-shot engine. The decisive experiment is one neither has run in public: take a set of decisions with computable answers, run both, and check whether the probabilities mean anything. Laya has half-published that experiment and shown you where it fails. Intern-Decision-4B has not published it at all.