ZwZ2-9B vs MathForm-8B: Same Trunk, Two Very Different Amounts of Truth

Two models published eight weeks apart, from two different labs, built on two different generations of the same family — and their trunks are the same width. MathForm-8B, released by OpenBMB on 14 August 2026 as a fine-tune of Qwen3-8B, runs 36 layers at hidden size 4,096 with an intermediate size of 12,288. ZwZ2-9B, posted by inclusionAI on 16 October 2026 on the Qwen3.5-9B stack, runs 32 layers at hidden size 4,096 with an intermediate size of 12,288. Parent generations apart, vocabularies of 151,936 and 248,320, entirely different jobs — and the same 4,096-by-12,288 trunk geometry in both. What separates them is not architecture. It is that one of them comes with a paper, a dataset, an evaluation harness and a licence, and the other comes with eight files.

The short version: if you can describe your problem as "turn this prose into Lean 4", MathForm-8B is a documented specialist with published numbers and you should use it. If your problem involves pixels and you want a general vision-language model, ZwZ2-9B is a 9.4B checkpoint nobody has characterised, and adopting it means becoming its first evaluator. The interesting part of the comparison is what that asymmetry teaches about how much a shared inheritance is worth.

Two descendants, one generation apart

MathForm-8B is a Qwen3-8B derivative and says so openly: its card records the base model, the training dataset, the training method and the paper. The config is a textbook Qwen3 layout — 36 layers, 32 query heads against 8 key-value heads at head dimension 128, a 151,936-token vocabulary, a 40,960-position ceiling, rotary base 1,000,000, and full attention in every layer. Nothing exotic, nothing hidden.

ZwZ2-9B is a Qwen3.5-9B derivative that says so only through a single Hub tag. Its config resolves to Qwen3_5ForConditionalGeneration, and every architectural value matches Qwen3.5-9B except one deleted multi-token-prediction layer. The 248,320-token vocabulary, the 262,144-position ceiling, the interleaved M-RoPE, and the vision tower are all inherited unchanged.

• Parameters — ZwZ2-9B: 9,409,813,744. MathForm-8B: 8 billion, per its card.

• Parent — Qwen3.5-9B vs Qwen3-8B, so the two inherit different attention designs from different Qwen generations.

• Attention — 24 linear-attention layers plus 8 full-attention layers vs full attention in all 36 layers.

• Trunk geometry — hidden size 4,096 and intermediate size 12,288 in both.

• Vocabulary — 248,320, with image and video token ids inside it, vs 151,936 and text only.

• Context ceiling — 262,144 positions vs 40,960.

• Modalities in — text, image and video vs text only.

• Evidence published — a config file vs a paper, a dataset, an evaluation pipeline and Apache-2.0 terms.

What MathForm-8B puts on the table

The card is unusually complete, and it is worth walking through because the contrast later depends on it. MathForm-8B translates natural-language mathematical statements into Lean 4. It was trained on FormalVerse — a dataset OpenBMB published alongside it — through supervised fine-tuning and then reinforcement learning, with Lean compilation and semantic-consistency feedback supplying the reward signal. The paper is MathForm: Scaling Mathematical Autoformalization with Knowledge Retrieval and Verification-Guided Refinement, arXiv 2608.14221, and the repository includes the evaluation pipeline, the benchmark files and the Pass@k scripts.

The results are reported as Pass@8 pass rates under two separate criteria: a Syntax Check, meaning the generated Lean actually compiles, and a Consistency Check, meaning it compiles and says the same thing as the original problem. Across six benchmarks the card reports a macro-average of 88.06 under syntax and 72.37 under consistency, ahead of the other 8B autoformalizers it lists and ahead of several 32B entries on the syntax column. That last comparison is the kind of claim worth testing rather than repeating, and there is a detail in the setup that makes testing it non-trivial: compilation checks require a running Kimina Lean Server, and the experiments used Lean 4.21.0. This is not a checkpoint you can score with a Python script.

Every number above is vendor-reported. OpenBMB ran the evaluation, OpenBMB published it, and the repository does include the scripts to reproduce it — which is more than most labs offer and is the reason the numbers are worth something even unverified. The card also publishes its recommended generation settings, temperature 0.6 and top-p 0.95, which is the sort of small operational detail that a model without a card leaves you to guess.

What ZwZ2-9B puts on the table

A config.json, a tokenizer_config.json, a chat template, and 18.8 GB of weights. There is no README, no licence field, no pipeline tag, and no dataset. The Hub reports zero downloads and two likes, and no inclusionAI collection contains it.

Reading the files does yield real information, and it is worth doing before dismissing the checkpoint. The chat template injects a default system prompt when a request arrives without one, and it defaults to not reasoning — it emits an empty <think></think> block unless a caller explicitly asks for thinking, which is the opposite of what its Qwen3.5-9B parent does. The config carries a 27-layer vision tower and a video processor accepting up to 768 frames at 2 frames per second. The parameters are 9,409,813,744 across 760 tensors, all bfloat16.

What none of that tells you is what the model is good at, what it was trained on, whether it is an instruction model or something else, or under what terms you may use it. MathForm-8B answers all four questions in its first screen. ZwZ2-9B answers none of them anywhere.

The shared 4,096 is not a coincidence, and it is not a shortcut

Both models use a trunk 4,096 wide with a 12,288 intermediate size, and the reason is the same in both cases: a Qwen generation that size. Qwen3-8B and Qwen3.5-9B are different designs — full attention throughout versus a three-to-one mix of linear and full attention — but they are both built on the same width, because that is what a small Qwen is. Watching a fine-tune inherit its parent's every dimension is the clearest evidence available that fine-tunes inherit their parent's properties wholesale: the strengths and the failure modes, not just the shape. MathForm-8B tells you which parent it came from, so you can go read about it. ZwZ2-9B does not, and the Hub tag is the only thread back.

Which one you actually need

These are not substitutes, and the overlap is close to zero. MathForm-8B does one thing: it turns a natural-language mathematical statement into Lean 4 that compiles. Its entire evaluation apparatus, including the requirement for a running Lean server, is built around that. If that is your problem it is a strong, documented, Apache-2.0 answer with reproducible tooling, and the 72.37 consistency figure is the number to beat — consistency, not syntax, because a formalisation that compiles but means something else is a bug that costs you more than a failed compile.

ZwZ2-9B is a general vision-language model. It has never been evaluated on autoformalisation, has no Lean tokenizer considerations to speak of, and was published by a lab whose earlier work was about fine-grained visual perception. Do not go looking for a ZwZ2-9B score on a Lean benchmark; there is no reason for one to exist.

The genuine question is whether you want a specialist or a generalist at all. A specialist is better at its one job and useless outside it. A generalist is worse at everything specific and is the only option when the input shape varies. The reason this comparison is lopsided is not the eight weeks between them or the billion parameters — it is that one of them was published with the documentation a specialist needs and the other was published as a starting point for someone else's work.

If you want both abilities in one pipeline

There is a real pattern where these two modes meet: a general model reads an ambiguous document, extracts the mathematical claim, hands it to a formaliser, and a checker decides whether the result stands. That is three model calls with different roles, and it is the shape that a routing layer exists to make cheap — OrcaRouter's routing DSL composes several models into a single call so a fan-out and a follow-up are one request rather than an orchestration service you maintain. Model fusion, where a panel of models answers together and a judge picks the strongest response, is the same idea applied to a single hard question rather than to a pipeline.

Neither ZwZ2-9B nor MathForm-8B is in our catalogue, and we are not going to dress that up: MathForm-8B is downloadable under Apache-2.0 and servable from your own vLLM or SGLang instance, and ZwZ2-9B is a repository with no endpoint at all. What is routable today from the reasoning side of that pipeline is qwen/qwen3.8-27b — Qwen's current-generation dense flagship, Apache-2.0, with a 262K-token context and a 1M-token hosted ceiling at $0.33 per million input tokens and $2.40 per million output, passed through at provider list price with nothing added.

What the asymmetry is worth

MathForm-8B is the better model to write about, because there is something to check. ZwZ2-9B is the more interesting artefact, because a 9.4B checkpoint with a tuned chat template and a stripped prediction head is evidence that somebody did real work — and because inclusionAI has a track record with the ZwZ name that makes the next post worth watching.

For anyone deciding today, the rule is unglamorous. A published paper, a dataset, a named base model and a stated licence are not paperwork. They are the difference between a model you can plan around and a model you have to reverse-engineer, and no amount of interesting architecture closes that gap.

A result like this is also a reminder that model mistakes are already documented: the AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.