Errata: GLM-5.3 is text-only — why it left the multimodal board

In the sixth edition we put Zhipu GLM-5.3 on the multimodal board at 95.4, tied for first with Qwen3.8-Max. In the seventh edition it is gone.

This was not a score revision, and the model did not get worse. We put it on the wrong board. GLM-5.3 is a text-only model: the official model card and API documentation list text as the only input modality. Image and video input belong to GLM-5.3-Flash. While mapping the Zhipu line-up, we attributed the family’s multimodal capability to GLM-5.3, then computed a composite from visual, video and chart scores that should never have existed for that entry.

What went wrong

The inclusion test for this board is simple: a model must accept at least one non-text input, and we must be able to score it across all eight dimensions. A text-only model is not “weak”; it simply has no perception, video or chart dimensions. Filling in zeros would bury it at the bottom, and estimating scores would invent readings nobody can verify. Both are dishonest.

The sixth edition did the second. Using public information about GLM-5.3-Flash, we estimated perception 97, documents 97, charts 96, video 93 and so on for GLM-5.3. Fed through the eight-dimension weights, those inputs produced 95.4 — a number that looks reasonable and describes the wrong product.

The top band after removal

Rank	Model	Composite	Evidence
#1	Qwen3.8-Max	95.4	B
#2	Kimi K3	95.3	A
#2	Doubao Seed 2.1 Pro	95.3	A
#4	Doubao Seed 2.0 Pro	95.2	A

The sixth edition had Qwen3.8-Max and GLM-5.3 tied at 95.4. With GLM-5.3 removed, Qwen3.8-Max is the sole leader, but the top four still sit inside the 0.5-point same-band threshold. The right reading is not “Qwen pulled ahead” but “the band lost a member; its width did not change.”

Kimi K3 and Doubao Seed 2.1 Pro tie for second at 95.3, ordered by raw score and evidence grade; Doubao Seed 2.0 Pro is fourth at 95.2. The largest gap in the top four is 0.2 points, so the next small update from any vendor can reorder them.

GLM-5.3-Flash is not a substitute

Why not let GLM-5.3-Flash inherit the slot, since it is multimodal?

Because they are different models with different parameters, context, licences and scores. GLM-5.3-Flash is a 320B total / 18B active open-weight model under MIT with a 1M context and a 93.7 composite, ranked #6; GLM-5.3 is a closed text flagship with a 400K context. Moving scores from one to the other would repeat the sixth-edition error in the opposite direction.

The sixth edition also mislabelled GLM-5.3-Flash as “closed, 200K, a distilled lightweight”. The seventh edition corrects it to MIT open weights, 320B-A18B and a 1M context. The score stays at 93.7, but the licence and deployment implications change completely: it can be downloaded, self-hosted and fine-tuned.

The other two removals

MiMo-V2.5. Xiaomi’s previous model was recorded in the sixth edition as a closed API with partial open components, with a 90.1 composite. It is replaced by MiMo-V2.6-Pro: 1T-A42B, 1M context, open weights, 91.4, #11. This is a generation change, not a retroactive mark-up.

InternVL 4.0. The “InternVL 4.0” entry in the sixth edition has no matching version in the official repository, model card or papers, so its parameters, licence and scores cannot be verified. We retired it and use two verifiable models instead: Intern-S2-Preview-397B (#26, 84.4, Apache-2.0) for scientific charts and InternVL-U 4B (#47, 75.8, Apache-2.0) for unified understanding and generation. They are not “new versions” of InternVL 4.0; they are two real models that can be checked.

What changed in the process

1. Modality gate first. Before inclusion we record the official input modalities. Text-only models do not enter the multimodal board and do not get estimated placeholder scores.
2. No inheritance inside a family. Different models from the same vendor and generation are documented separately; parameters, licence, context and modalities are checked one by one. Flash capability is never attributed to Pro.
3. Primary sources first. Parameters, licences and context come from official weight cards and API documentation; technical reports and community reproduction only supplement score evidence.
4. Removals are logged. Every removal, replacement and correction goes into the changelog. Old scores are not quietly overwritten.

What did not change

• The scorecard, eight weights and 0.5-point same-band rule are unchanged.
• Apart from the 24 entries added or replaced this cycle, the other 32 surviving entries were not re-scored because of this removal; rank changes come from the new sample and the top band.
• Essays from the sixth edition and earlier keep their original sample (35 models / 280 points) and now carry a sample-period note. Their GLM-5.3 ranks and scores are historical records, not the current board.
• Sample-dependent derived statistics still use the sixth-edition basis and will be recomputed across 56 models next.

What the error cost

It did not change the tier-level conclusion about who leads — the top four were already in one band. It did change one concrete procurement recommendation: any team that chose GLM-5.3 for image or video work because it “tied for first on a multimodal board” would have selected the wrong product. That is the real cost of a classification error.

A board can carry a judgement call that later changes. It cannot carry a score nobody can trace. Removal, correction and a public record are worth more than a flattering number.