GLM-Legal vs TwIL-LM3-Pro: The Gate Is a Licence on One Side and a Form on the Other

A legal agent that exists as a benchmark score and an application form, against a formal-logic model you can download this afternoon and deploy almost nowhere. GLM-Legal and TwIL-LM3-Pro are both narrow specialists released weeks apart — one built by Z.ai on GLM-5.3-Flash for contract review, evidence analysis and compliance, the other a 3.66B-parameter reasoning model from webAI built for formal logic and derived from IBM's granite-4.2-3b. They are not rivals for the same slot, and this page is not going to pretend otherwise. What makes them worth putting side by side is that they are the two distinct shapes a specialised model takes in 2026, and each shape has its own way of being unavailable: TwIL-LM3-Pro ships under a licence that forbids commercial use, and GLM-Legal ships behind a beta application with no published licence, price or endpoint at all.

If you arrived here deciding which to adopt, the short answer is that one is a file you cannot legally use and the other is a service you cannot yet buy. The longer answer is about which of those two problems is actually solvable this quarter.

The specs, and what they rule out

The parameter difference is large enough that it changes what each model is for, so it is worth putting the numbers down before the argument:

• Size and shape — GLM-Legal is a derivative of a 320B-total, 18B-active multimodal foundation model; TwIL-LM3-Pro is a 3.66B dense decoder, 3,659,737,600 parameters exactly, 40 layers, and it runs in bf16 at 6.82 GiB or as a Q4_K_M GGUF at 2.09 GiB.

• Context — GLM-Legal inherits GLM-5.3-Flash's 1M-token window; TwIL-LM3-Pro carries 131,072 tokens from granite-4.2-3b unchanged.

• What it was trained on — legal corpora in mid-training plus RL across tens of thousands of realistic legal environments vs a rank-64 LoRA supervised fine-tune on a synthetic formal-logic corpus, multipath distillation, checkpoint fusion, and an entropy-weighted GRPO stage against a programmatic verifier.

• Modality — GLM-Legal's base takes text, image and video in; TwIL-LM3-Pro is text in, text out, with a <think>…</think> block before the answer.

• Distribution — an application form for API access vs a Hugging Face checkpoint, a LiteRT-LM on-device build, and an enterprise arrangement.

• Published benchmarks — one all-pass number on a third party's harness, plus two unquantified parity claims, vs a detailed set of vendor-run numbers on a vendor-built harness that the model card itself describes as inconclusive in places.

Those two rows at the bottom are the ones that decide everything else. Read the deployment row first: GLM-Legal has no public endpoint and TwIL-LM3-Pro has no commercial licence. Everything else in this comparison is downstream of that.

Reading TwIL-LM3-Pro's numbers the way its own card does

webAI published unusually candid evaluation material, and the most useful thing in it is a sentence that would have been deleted from a marketing page. On the macro gate — 0.5539 against Qwen3-8B's 0.5336 — the card writes that the lead is 0.020, "smaller than the sampling noise at n = 200 per lane," and tells the reader to score it as parity rather than a win.

Where the model does separate from Qwen3-8B, the margins are real and measurable:

• Strict-7 — 0.2879 against 0.2093, a row that uses no loose-match credit anywhere and so cannot be produced by formatting luck.

• Strict MCQ — 0.4100 against 0.0000, the widest lane gap in the eleven reported rows.

• Corpus fit — the lowest language-model perplexity of any arm at 2.3130, against 2.5478.

• Held-out suite — the model does not lead it. The ten-dataset macro is 0.7901, statistically level with its own base at 0.7942 and behind Qwen3-8B's 0.8493.

• Inside the specialisation — FOL translation exact match sits at 0.0100 and semantic-parse exact match at 0.0000, both disclosed by the vendor.

That is what a narrow specialist looks like when it is documented honestly: strong on the exact lanes it was built for, level or behind on general reasoning, and poor on two specific output formats. It is also the reason the licence matters so much. A model this specific has one job, and if that job is commercial, the licence blocks it entirely.

Reading GLM-Legal's numbers with the same suspicion

Z.ai's release leads with 19.2% all-pass on Harvey LAB, a law-firm workflow evaluation across 24 practice areas, and a 12.5-percentage-point improvement over GLM-5.3-Flash. Both are the vendor's own figures. The second one implies a base-model all-pass rate somewhere around 6.7%, which tells you how harsh the all-pass standard is: a task counts only when every step lands, so a low absolute number is the expected result of strict grading rather than evidence of a weak model.

The two other headline benchmarks carry no numbers. LegalBench — contract parsing, merger filings, judicial opinions — is described as matching "the precision of top closed-source frontier models." PLawBench, the Chinese-practice suite, is described as reaching "parity with leading frontier systems." Neither claim names a comparator or attaches a score. Applied to TwIL-LM3-Pro's card, the same standard asks how the legal model does on the format-exactness lanes, and there is no answer, because no one has published one.

The one thing GLM-Legal publishes that TwIL-LM3-Pro does not is a demonstration of its specialisation working. On a 40-document project dispute, GLM-5.3-Flash finds the contradiction between a 60%-complete client report and a 41% internal snapshot and says the documents may support the client's position. GLM-Legal links an earlier internal warning, the snapshot and an instruction to suppress the risk and keep invoicing into one concealment theory, then cites the agreement's payment-does-not-waive-rights clause and names the claim elements still to test. Whether that generalises is unknown. That it is a materially different behaviour from its base is visible.

The honest answer to "can I use this in production"

Both answers are no, and the two noes have different expiry dates.

TwIL-LM3-Pro's blocker is a licence, and licences change with a decision. The checkpoint sits on Hugging Face under webAI's own non-commercial terms; a relicense or a commercially licensed LiteRT-LM build would move the model from research artifact to deployable component in a single afternoon. The on-device build that appeared on 1 October carries the parent repository's restriction, which is the thing to verify before planning around it.

GLM-Legal's blocker is a programme, and programmes change with a contract. Access runs through an application aimed at law firms, legal-tech companies, in-house teams and independent practitioners, with no published price and no public endpoint. The AutoClaw integration is written in the future tense. For a regulated deployment, this is a harder problem than a restrictive licence, because you cannot build on something you cannot call.

There is also a data question that neither release answers. webAI says the family passed half a million downloads in a month — a vendor figure, unaudited — and downloads of a free non-commercial checkpoint are not a proxy for production use. GLM-Legal's beta terms are unpublished, which is different from being unacceptable and identical to being unassessable by a risk committee.

What to build on while both gates are shut

The practical move is to separate the specialist from the pipeline, because the pipeline is what you will actually own. Neither specialist is routable anywhere today; what is routable is the generalist underneath GLM-Legal and the surrounding steps.

GLM 5.3 Flash is on the OrcaRouter catalogue at Z.ai's list price with 0% markup — $0.15 per million input tokens and $0.50 per million output, cache reads at $0.030 — so a vendor price change on the base model is live on our side the same day. One key reaches more than 200 models, which is what lets a legal engineering team prototype the drafting, templating and checklist steps on the base model now and change a single model id when the beta opens. Automatic failover makes that prototyping safe: a staging route can point at a model nobody has independently evaluated without putting a production path behind it.

For the formal-logic half of the picture, the local checkpoint and the hosted generalist are complements rather than substitutes, and model fusion is the mechanism that makes that concrete — a panel of models answering together, so a small runner doing strict entailment or Lean statement formalisation can sit alongside a hosted model doing the prose around it. The split webAI's own numbers suggest is the obvious one: run the local model where the deliverable is a label, and route everything that has to come back as writing.

The one thing not to do is treat either of these as a decision. TwIL-LM3-Pro is a 2.09 GiB download that costs nothing to evaluate and nothing to abandon. GLM-Legal is a form, and filling it in early costs you an afternoon.

Which one, if it is genuinely one or the other

They do not overlap enough for a single answer to be useful, so the honest version is conditional. If your problem is entailment labelling, Lean statement formalisation, semantic parsing, or an air-gapped pass where the output is a label rather than prose, TwIL-LM3-Pro is the closer fit of the two and the licence is your only real obstacle. If your problem is drafting to a specification, statutory compliance across named regimes, or reasoned analysis over a document set you already hold, GLM-Legal is aimed at it and TwIL-LM3-Pro has published no legal evidence at all.

What they share is that neither vendor has put a price on the specialised product, and both have graded their headline result against their own base model on their own harness. That is not a reason to dismiss either release — GLM-Legal's 19.2% is credible precisely because the grading is harsh, and TwIL-LM3-Pro's card is honest enough to print its own null results. It is a reason to plan the integration around a model you can call today and treat both of these as upgrades you adopt when the gate opens, not as foundations you build on.

The gate on TwIL-LM3-Pro is legal text, and legal text is the one thing in this comparison that a vendor can change without shipping anything. The gate on GLM-Legal is a business decision about when a beta becomes a product. Of the two, watch the licence — it is the one that could move without warning.

model fusion is the mechanism that makes that concrete — a panel of models answering together, so a small runner doing strict entailment or Lean statement formalisation can sit alongside a hosted model doing the prose around it.