Naming archaeology: the 32-model naming system and the three things a name actually predicts

A name is a model’s first spec sheet, and the only one that needs no third-party reproduction, no evidence grade, and can be read by everyone before the launch event.

32 China-native multimodal models use 32 names. The names pack in brand (Qwen, GLM, ERNIE), generation (3.5, 5.3, V4.1), size (Max, Pro), speed (Flash), modality (Omni, VL, Vision) and pure codenames (K3, N1, X2, Seed). This essay splits all 32 names into tokens and answers one question: what does each word in a name actually foretell — capability, or price.

The conclusion is not complicated: a name foretells price far more accurately than capability; and the loudest words in a name (Max, Pro, Omni) promise the least about capability direction. This is an angle nobody has written, because it needs not benchmarks but a linguistic inventory.

Five naming strategies, one comparison table

Split the 32 models into five families by the most prominent word in the name (each model counts once, priority: size word > speed word > modality word > generation word > no word):

Naming family	Representative word	Models	Average composite	Average output price	Average modality count
Size word	Max / Pro	3	93.5	¥15.0	4.0
Speed word	Flash	2	87.7	¥0.9	3.5
Modality word	Omni / VL / Vision	5	84.1	¥3.8	3.6
Generation word	Next / Preview	1	89.2	¥2.5	4.0
No word	Pure codename	21	83.6	¥7.7	3.1
Site-wide	—	32	85.0	—	3.3

Read the three columns side by side and the table speaks for itself: the size-word family (¥15.0) and the speed-word family (¥0.9) differ 16.7× in price, but only 5.8 points in composite (93.5 / 87.7). In score ratio, a name discriminates price by about 17× and capability by about 1.07×.

The more telling row is the modality-word family: the five models with Omni / VL / Vision in the name average 84.1, below the site-wide average of 85.0. A model whose name says “all modalities” is not stronger than average on the overall board.

Finding one: size words are price labels, not capability labels

“Max” and “Pro” mean “the strongest one” in launch-event language. The data supports only half of that:

• Qwen3.8-Max 95.4, #1 overall, ¥18;
• Doubao Seed 2.0 Pro 95.2, #2 overall, ¥15;
• SenseNova 6.5 Pro 89.9, #8 overall, ¥12.

Two of the three sit in the top two overall, so “Pro/Max ranks higher” is not wrong. But add price and the rule deforms immediately: the three average the most expensive price tier on the site (¥15.0), and the average score they buy (93.5) is only 5.8 points above the speed-word family. A size word foretells “you will pay more”, not “you will score more”.

In reverse, the speed-word family has the lowest price (¥0.9) and the second-highest composite (87.7). DeepSeek-V4.1-Flash and Qwen3.8-Flash prove one thing: in 2026, the word “Flash” foretells not “weaker” but “cheaper”.

Finding two: modality words are neither necessary nor sufficient

“Omni” is the strongest word semantically and the weakest to deliver in the name inventory. 13 models on the site have modality count ≥4, and only 2 of them carry Omni in the name (Qwen3-Omni-30B, WITA-Omni); conversely, of the 5-modality models at the top of the modality count (Qwen3.8-Max, WITA-Omni), half have no modality word in the name at all.

The more direct counterexample is Keye-VL-2.0 and Hunyuan-Vision: the names say VL and Vision, yet both have modality count 3 — narrower than Kimi K3 (4 modalities) and MiniMax M2.5 (4 modalities), which say nothing in the name.

So the correct reading of a modality word is: it does not guarantee multimodality; it only guarantees that the product line put “vision” on the roadmap. VL / Vision in the name means the vendor is willing to sell vision; it does not mean the modality coverage will be wider than average.

Finding three: the family prefix signals ecosystem, the suffix signals price

Four Qwen models are the cleanest sample of this rule:

Name	Suffix	Output price	Ecosystem score	Composite
Qwen3.8-Max	Max	¥18	96	95.4
Qwen3.5-397B-A17B	397B	¥12	98	94.6
Qwen3-Omni-30B	Omni	¥4	90	88.9
Qwen3.8-Flash	Flash	¥1.2	97	87.5

The prefix is the same (Qwen), price spans 15× (¥1.2–18), and the ecosystem scores all land between 90 and 98. The prefix tells you “this is the same toolchain”; the suffix tells you “which price tier you are paying for”. Ecosystem score is almost unrelated to the suffix and strongly related to price and positioning.

The rule has a direct use for selection: when you pick a model inside one family, do not sort by “which name sounds stronger” — sort suffixes by price band, then compare capability inside that band. DeepSeek-V4.1-Flash is the value anchor of October 2026 precisely because it put “document 95” behind a Flash suffix.

Finding four: only two models put parameter counts in the name

Of the 32 names, only 2 write a parameter spec into the name: Qwen3.5-397B-A17B (397B total / 17B active) and Qwen3-Omni-30B (30B-A3B). The other 30 names carry only codenames, generations and adjectives.

This is two sides of the same fact as the parameter transparency index: of the 32 models, only 5 publish parameter counts, and only 2 are willing to print them in the name. The numbers in a name (3.5, 5.3, V4.1) are version numbers, not specs; reading a version number as a capability scale is the most common illusion in naming archaeology.

Version numbers are valid only within a family: Qwen3.8 > Qwen3.5 (95.4 > 94.6), GLM-5.3 > GLM-5 (92.4 > 92.0), both generational upgrades hold. Across families they do not: Emu3.5 (3.5) ranks #24, Kimi K3 (integer 3) ranks #5. A version number is a coordinate inside a family, not a capability scale across the site.

Procurement reading: use the name as an index, not as evidence

This essay has one usable conclusion: a name can narrow the candidate set, not sort it.

1. Filter ecosystem by prefix: for the same suffix (say Flash) across Qwen and DeepSeek, toolchain, pricing model and open-source licence are the main differences; capability differences are secondary.
2. Filter budget band by suffix: Max/Pro is the most expensive tier, Flash the cheap tier, Omni the widest-priced tier (¥0.5–8.6) — fix the budget first, then read capability.
3. Ignore modality words: for modality coverage, consult the modality negative-space map directly, which holds the real bundling of all 32 models, instead of the adjectives in the name.
4. Do not compare version numbers across families: comparing “who is 3.5 and who is 5.3” across families is a pure pseudo-ranking.

Conversely, if your selection process contains reasoning like “it has Pro in the name, so it is stronger”, this naming-archaeology essay was written for that step.

Boundaries

• Naming and capability are not entirely unrelated. The size-word family is indeed the highest-averaging family on the site, so “the name has no information” is a wrong conclusion; the correct statement is “the name’s information falls mainly on price and positioning”.
• The classification is single-label and loses information. A model can carry a modality word and a generation word at once (WITA-Omni Preview); this table takes only one by priority; change the priority and the family averages move slightly.
• The sample is 32 models, at 2026-10-18. Naming strategy changes as vendors adjust roadmaps; trends like “Flash getting stronger” are exactly the quantity this essay wants to monitor.
• Averages hide distribution. Within the modality-word family, WITA-Omni 80.6 and Qwen3-Omni-30B 88.9 differ by 8 points; the family average 84.1 is a flattened number.

Reproduction

Every input to this table lives in the site’s model data: names from rankName, prices from priceOut, modality counts from the length of the modalities array, and composites recomputed with the site weights. The family-splitting rule is stated in the body; you can rerun it with a different priority.

If this essay leaves one sentence: a name foretells price, not capability; filter candidates by name, make decisions by the dimension scores in the archive.