Intern-Decision-0.8B vs MathForm 8B: One of These Can Check Its Own Work

The most useful question you can ask about any small specialist model is who checks it. Intern-Decision-0.8B and MathForm 8B both exist because a general model was fine-tuned into a narrow one, both are Apache-2.0 derivatives of Qwe​n weights, and both were put on Hugging Face without a marketing campaign — but they sit on opposite sides of a line that determines how you would deploy them. MathForm 8B is OpenBMB's 8B autoformalization model, released on 14 August 2026, that translates natural-language mathematics into Lean 4 and hands the result to a compiler, which either accepts it or does not. Intern-Decision-0.8B is InternLM's 852,985,920-parameter decision head, uploaded on 26 September 2026, that returns a calibrated probability distribution over options you supplied in advance — and nothing in the world independently verifies whether the label it returns is right. One of these models produces output with a proof. The other produces output with a confidence score, and the difference in what you can do with those two things is the entire article.

That framing also explains why the size gap — 8B against 0.8B, a factor of ten — is the least interesting number in the comparison. Neither model is trying to be good at what the other does, and neither one's evaluation tells you anything about the other. What they share is a release pattern and a Qwen lineage, and both of those matter less than the verification question.

What each one actually is

MathForm 8B is the shipped artifact of a two-stage recipe described in the paper MathForm: Scaling Mathematical Autoformalization with Knowledge Retrieval and Verification-Guided Refinement (arXiv 2608.14221). A retrieval planner pulls relevant definitions and existing formalizations from Mathlib before the generator runs; generated statements are then revised using compiler diagnostics and semantic-consistency feedback; the resulting corpus, FormalVerse, contains approximately 367,000 verified Lean 4 examples; and MathForm 8B is trained on it through supervised fine-tuning followed by reinforcement learning using Lean compilation and semantic-consistency signals as the reward. It is a text-only model based on Qwen3-8B, served through Transformers, vLLM or SGLang with an OpenAI-compatible API, with a recommended 16,384-token context length and a 16,384-token maximum generation budget. Its evaluation pipeline, benchmark files and Pass@k scripts are all in the OpenBMB GitHub repository, and the training dataset is public.

Intern-Decision-0.8B is the shipped artifact of a recipe nobody has described. The model card says it is "a multimodal structured decision model fine-tuned from Qwen3.5-0.8B" and stops there — no data description, no training procedure, no paper, no repository. What it does document thoroughly is the inference contract. The bundled engine maps each question's options to single-token symbols, renders a skeleton with one <decision> placeholder per field, runs one causal forward pass, reads logits at the position before each placeholder, softmaxes over only that field's permitted symbols, applies a fitted calibration, and maps symbols back to your option values. There is no generate() call and no sampling anywhere in the path. It accepts text plus up to eight images, handles one to sixteen questions with up to 62 options each, and rejects inputs past 8,192 tokens rather than truncating them. Default calibration temperature is 2.747760550703, fitted per checkpoint by NLL minimisation over 1,728 cases.

Read those two paragraphs side by side and the asymmetry is stark. MathForm 8B comes with a paper, a dataset, an evaluation pipeline and a repository. Intern-Decision-0.8B comes with an API description and a benchmark table.

The verification asymmetry, which is the real story

MathForm 8B's output is checkable by something other than a human. It emits Lean 4, and Lean 4 either compiles or it does not. OpenBMB's numbers are reported under two regimes for exactly this reason: Pass@8 of 88.06% under Syntax Check, which means the output compiles, and 72.37% under Consistency Check, which means it compiles and a semantic-consistency check agrees the formal statement means what the informal one said. Both figures are averages across six benchmarks, and the paper also reports CC pass rates of 63% on FATE-H and 37% on the harder FATE-X subsets, with the claim that this exceeds specialised 32B autoformalizers. Those are vendor-reported — OpenBMB ran them — but the property that matters is structural, not statistical: a downstream system consuming MathForm 8B's output can reject a bad formalization without asking a model to judge it. The compiler is the oracle.

Intern-Decision-0.8B has a type contract and no oracle. The output shape is guaranteed — a field declared choice returns a distribution over the option values you listed, a score returns a probability-weighted expected value over your rubric, a noul returns a yes-probability. Nothing outside the model can tell you whether the argmax was correct. The confidence value is the model's own estimate of its own correctness, and the card's calibration work is an honest attempt to make that estimate meaningful — a fitted temperature that preserves the argmax while sharpening or softening the probabilities, validated on held-out cases — but a well-calibrated wrong answer is still a wrong answer. If your pipeline needs to know whether a label is right, you need labelled data, and you need to measure it yourself.

That is not a defect unique to Intern-Decision-0.8B. It is the condition of every classifier, and it is the unresolved problem in the whole category of decision models that includes TypeSafe's Jev and Convai's Laya. It is worth stating clearly because a benchmark table with a Brier score column can create the impression that calibration is verification. It is not. Calibration tells you that when this model says 80%, it is right about 80% of the time across the evaluated distribution — which is genuinely useful for setting thresholds and computing expected value, and is not a per-item correctness guarantee.

The scoreboard, on the rows both models have

• Parameters — Intern-Decision-0.8B: 852,985,920 across a 1.50 GB language shard, a 176 MB vision shard and a 25 MB projector. MathForm 8B: 8B dense, based on Qwen3-8B.

• Base model — Intern-Decision-0.8B: Qwen3.5-0.8B, released February 2026. MathForm 8B: Qwen3-8B.

• Task — Intern-Decision-0.8B: typed decisions over a schema you write — choice, score, binary. MathForm 8B: natural-language mathematics to Lean 4 formalization.

• Output — Intern-Decision-0.8B: calibrated distribution plus argmax per field, no text generated. MathForm 8B: generated Lean 4 source, typically long.

• Verification — Intern-Decision-0.8B: none external; confidence is self-reported. MathForm 8B: the Lean 4 compiler, plus a semantic-consistency check.

• Published evidence — Intern-Decision-0.8B: a vendor benchmark table across seven benchmarks, unreproduced, no paper. MathForm 8B: a paper, a public dataset of roughly 367,000 examples, an evaluation pipeline and repository, vendor-run.

• Licence — Both Apache 2.0, and Intern-Decision-0.8B additionally carries the preserved Qwen licence file for its upstream weights.

Cost and latency are not comparable, and that is not a dodge

InternLM measured Intern-Decision-0.8B at 33.98 ms mean and 37.50 ms p95 per query on a single RTX 4090 through the local Hugging Face path, with its 2B sibling posting 33.28 ms mean. MathForm 8B's own card recommends up to 16,384 new tokens per formalization at temperature 0.6 and top_p 0.95. Those two measurements are not the same quantity. One is a single forward pass over a prompt; the other is an autoregressive generation that can run for thousands of tokens. Multiplying out a formalization budget against a 34 ms decision tells you nothing about relative efficiency, because the models are not doing the same amount of work — one reads and scores, the other reads and writes a proof script. If throughput is your constraint, the relevant facts are simpler than a ratio. MathForm 8B formalizes one statement per generation, and at 16K tokens per output on an 8B dense model, that is a GPU-saturating workload and a candidate for batching through vLLM or SGLang, both of which OpenBMB documents. Intern-Decision-0.8B answers sixteen questions covering an entire record in one pass, so the unit of work is a record rather than a field, and the record budget is the 8,192-token input ceiling — which arrives faster than a reader might expect when you are packing in a long state, a rich schema and up to eight images.

Where each one is the right tool

MathForm 8B belongs in a pipeline whose bottleneck is human review of mathematics. Autoformalization exists because writing Lean is slower than reading it, and because a machine-checkable statement is one a proof assistant can then attack. The property that makes it trustworthy — compiler-verified output — is also the property that makes it narrow: it formalizes, it does not prove, and the card is explicit that compilation checks require a running Kimina Lean Server and that the experiments used Lean 4.21.0. Anyone adopting it is adopting that stack. As with any days-old or weeks-old open-weights model, pointing a test path at it and falling back to a proven model when it stalls is the low-risk way to evaluate it, which is what a gateway with automatic failover across a fallback chain is for — retries that land before the response starts, so a stalled formalization never reaches your caller.

Intern-Decision-0.8B belongs wherever a closed answer set already exists and the cost of generating text to recover it is pure waste. Triage, routing, rubric scoring, adjudicating a record against a written policy — the cases where a generative model is being used as an expensive way to pick from a list. Its advantages are that it is deterministic, that it returns a usable probability rather than a string you have to parse, and that at 1.73 GB on disk it runs comfortably under the memory cost of a laptop browser. Its disadvantages are that the documentation stops at the API surface, that no one outside InternLM has published a result for it, and that the one benchmark column suggesting a safety problem — a WildJailBreak score of 64.48 against Jev's 96.29 — is unexplained. Do not put it in front of adversarial input without running that test yourself.

Both models share a lineage worth noting because it changes what "open" buys you. Each is a fine-tune of a Qwen checkpoint, and each preserves the upstream licence correctly: MathForm 8B under Apache 2.0 with the Qwen3-8B provenance stated on the card, Intern-Decision-0.8B under Apache 2.0 with a separate LICENSE-QWEN file in the repository. Neither carries a revenue threshold or a field-of-use restriction — unlike Liquid AI's LFM Open License v1.0, which conditions commercial rights on your entity staying under a $10 million annual revenue line. If you are building commercially, that is a distinction that separates these two releases from parts of the small-model ecosystem, and it applies to both equally.

If you want the decision layer without the undocumented checkpoint

The gap between "a decision head is the right shape for this problem" and "this particular decision head is one you can defend to a reviewer" is what a hosted alternative closes. TypeSafe's Jev 1.13 is the model InternLM benchmarked its family against on Jevbench, on Typed Decision and on ToolACE, and it is callable today through a single OpenAI-compatible endpoint at $0.042 per million input tokens with output billed at zero — the vendor's published rate, passed through with no markup rather than marked up on the way. For a reader who wants to measure whether a decision head helps at all before committing to a 0.8B checkpoint with a missing repository, that is the cheap first experiment, and Jev's own third-party evaluations give it a documented track record that Intern-Decision-0.8B does not yet have.

The short answer

These two are not alternatives. MathForm 8B is a specialist whose output a program can verify, aimed at a task where verification is the hard part, and it comes with the paper, dataset and evaluation harness to prove the claim. Intern-Decision-0.8B is a specialist whose output only you can verify, aimed at tasks where the answers were already written down and the hard part was getting to them quickly and cheaply, and it comes with an inference module and a table. If you need a formalization, there is only one of these to consider. If you need a label and you are prepared to build the ground truth to check it against, the 0.8B is the more interesting download — fast, deterministic, Apache 2.0, and small enough that the cost of finding out whether it is any good is an afternoon rather than a budget line.