Intern-Decision-2B vs Laya: 2.2 Billion Parameters Against 421 Million, Both Refusing to Write an Answer

internlm/Intern-Decision-2B and convaiinnovations/laya are the two most similar models in the small decision-model niche, and the most useful fact about the comparison is that they get there by opposite routes. InternLM's 2,213,241,664-parameter checkpoint reads the logit at a fixed position before a placeholder and never calls generate(). Convai's 421-million-parameter checkpoint feeds typed questions to a decision head on top of a ModernBERT-large backbone and returns calibrated probabilities in a single forward pass. Neither writes a token, both promise probabilities, both cap out near eight thousand tokens of input, and the smaller one is five times smaller and roughly a thousand times more popular. What separates them is not capability so much as packaging, and the packaging gap decides the purchase for most readers.

Intern-Decision-2B appeared on Hugging Face at 05:36 UTC on September 26, 2026, the middle of three sizes InternLM uploaded in forty seconds with no announcement, fine-tuned from Qwen/Qwen3.5-2B under Apache-2.0. Laya was first pushed on September 18, 2026 and has since accumulated 3,700-odd likes, a pip package, a Jev-compatible HTTP server, ONNX and LangChain integrations and a fine-tuning notebook. The contrast in maturity is the article.

The premise they share, and the mechanisms that differ

Both cards argue that if your problem is choosing among known answers, generating prose is the wrong operation. Laya's phrasing is "it never generates text, so there is nothing to parse and nothing to hallucinate." Intern-Decision-2B's is that its API "performs structured candidate scoring. It does not call generate() or sample free-form text."

The mechanisms are where they part.

Laya is a classifier. A ModernBERT-large encoder (395M, bidirectional, fully fine-tuned) feeds a decision head trained from scratch — two transformer layers, an option-marker scorer, and an act/escalate head — for 421M total. Options share a fixed token budget (head_max_len, 192 tokens on the English checkpoint, 256 on multilingual), and the router detects script and language in under half a millisecond before the pass.

Intern-Decision-2B is a next-token model that has been forbidden from becoming a generator. It maps each question's options onto single-token symbols A–Z, a–z, 0–9; renders a full assistant JSON skeleton with one <decision> placeholder per field; runs one causal forward pass; reads logits immediately before each placeholder; softmaxes over that field's legal symbols; and applies a fitted temperature of 2.100509348278. Underneath is a standard Qwen3_5ForConditionalGeneration — 24 layers, three linear-attention layers to one full-attention layer, a retained multi-token-prediction layer, a 262,144-position embedding ceiling — plus a vision tower and projector.

The practical consequence of the difference is option capacity. Laya's fixed per-option budget collapses on wide label spaces; its own card documents a 77-label question scoring 0.425 against TypeSafe Jev's 0.870 on the same task, because 77 options receive about three or four tokens each. Intern-Decision-2B takes up to 62 options per question and up to sixteen questions, and 62 is not a preference but the exact count of single-token symbols its contract can address. Neither is comfortable past a few dozen options; the failure shapes differ.

Scoreboard

• Parameters — Intern-Decision-2B 2,213,241,664 in BF16 plus a vision tower and projector, about 4.46 GB on disk; Laya 421M, a single 842 MB safetensors file, with a separate 644 MB multilingual checkpoint.

• Input ceiling — 8,192 tokens for both, both refusing rather than truncating; note that Laya's 8,192 applies to laya-multilingual and requires max_len=8192, since the shipped default is 1,024.

• Images — up to eight for Intern-Decision-2B, counted against the same token budget; no multimodal path for Laya.

• Speed — 33.28 ms mean, 33.15 ms P50, 33.55 ms P95 per request on one RTX 4090 for Intern-Decision-2B; Laya reports roughly 33 ms per request and 103–332 questions per second batched on a single T4.

• Languages — Intern-Decision-2B makes no multilingual claim at all; Laya routes across 100+ languages, reporting 45 of 51 benchmarked languages usable at over three times random, and 0.451 on MASSIVE intent across thirteen non-English languages against 0.306 for its English-only checkpoint.

• Zero-shot accuracy — Intern-Decision-2B posts 84.68 averaged over seven vendor suites with Brier 0.437 and ECE 0.100, all unreproduced; Laya's base English checkpoint scores 0.362 on the 2,000-decision typed-decisions benchmark, which its own card flags as below the 0.461 majority-class line.

• Packaging — a checkpoint, an inference.py and a newly public GitHub repository with training and evaluation code, versus pip install laya, a Jev-compatible HTTP server, ONNX Runtime and TileLang fast paths, MCP and LangChain integrations, and a fine-tuning notebook that runs on free tier GPUs.

The fairest comparison is not the one the spec sheet sets up

Putting 84.68 beside 0.362 is close to dishonest, and it is worth saying why rather than letting the numbers stand.

Laya's 0.362 is a base checkpoint measured zero-shot on a benchmark it was not tuned for. Its own card draws the conclusion explicitly: "Laya is a fast base to specialise, not a zero-shot decision engine." Fine-tuned on that benchmark's own training split it reaches 0.766, clearing the 0.735 teacher ceiling and beating TypeSafe Jev's published 0.727 on the same decisions. Intern-Decision-2B's 84.68, meanwhile, is a seven-suite vendor average whose test sets are not the ones Laya quotes, produced by the lab that built the model, with no third party having run it — the checkpoint shows one like and zero downloads. Comparing a fine-tuned result to a zero-shot result, or a vendor average to an independent board, is not a comparison. It is two different measurements sharing a typeface.

What is genuinely comparable: both are small, both are cheap to run, and both are unfine-tunable into your domain. The repository InternLM published this morning makes that last part materially easier than it was yesterday, because it ships the training launcher, the masked-next-token objective, a hash-verified bundle of 10,751 test rows across seven suites, and the temperature-fitting and replay scripts. A lab that ships a deterministic calibration generator and asserts replay changes zero decisions is inviting exactly the kind of adaptation Laya's card recommends.

How you would actually call either one

Neither model is on OrcaRouter. OrcaRouter's model page for Intern-Decision-2B returns 404 and there is no Laya route either; nothing here is an availability claim. Intern-Decision-2B means downloading the checkpoint and running its bundled engine on your own GPU. Laya means pip install laya and, if you want the Jev-compatible surface, laya-serve on POST /v1/systemone.

The Orchestrator-shaped question underneath is whether you want a second operational surface for a scorer. Laya is designed to be embedded — the same request and response shape as TypeSafe Jev, so an existing client changes a base URL and nothing else. Intern-Decision-2B is designed to be reproduced, and today it is roughly a day of environment work before the first decision comes back. If the closed-set decision you are making is close in shape to either of these, the hosted alternative that both cards benchmark against is TypeSafe Jev 1.13, which does have a route: $0.042 per million input tokens, a 65K context and a P50 time-to-first-token of 178 ms, reachable on the same key as 200-plus other models with provider list price passed through at 0% markup.

Where each one wins

Laya wins on everything operational. A pip package against a checkpoint download. A router across more than a hundred languages against no multilingual claim. Batched throughput measured in hundreds of questions per second against a per-request figure with no batching story. Two years of ecosystem muscle — ONNX, TileLang, LangChain, MCP — against a repository that has been public for two hours.

Intern-Decision-2B wins on three narrow things. It takes images, which Laya does not. It carries no fine-tuning debt: its 84.68 is a zero-shot vendor claim, where Laya's headline 0.766 belongs to a checkpoint fine-tuned on the benchmark's own training split. And it is documented with a reproduction kit — a verifiable evaluation bundle and a public training stack — which is worth more than a leaderboard row to anyone who has to justify the model to someone else.

The tie is the premise. Both refuse to generate, both give you probabilities you can threshold, and both are below the input length at which most real documents live. If your state is a ticket, an email or a form, either will do and Laya will get you there faster. If your state has a screenshot attached or your organisation needs the reproduction to be auditable, InternLM's middle checkpoint is the one that answers that specific objection — and on InternLM's own numbers it is the least well-calibrated of its three siblings, at ECE 0.100 against 0.066 and 0.065, so fit your own temperature before you trust the confidence it reports.