The models that do not exist: a negative-space map of 32 Chinese multimodal models, and four combinations nobody ships

First what exists, then what does not

Multimodal reviews usually list supported modalities: text, image, video, audio, GUI, embodied. More ticks look stronger.

But a tick list has a blind spot: it shows combinations that exist, not combinations nobody ships. Across 32 Chinese models, the modality configurations are not 32 different shapes; they are eight bundles repeated 32 times. Put those eight bundles into the 64 possible subsets of six modalities and the empty space is large.

This essay maps the negative space: list the eight bundles that exist, then four combinations nobody ships, then answer the more useful question — which gaps are real demand and which are merely unnecessary.

32 models, eight modality bundles

Using our six-modality taxonomy (text, image, video, audio, GUI, embodied), coverage is:

Modality	Models	Share
Text	32	100%
Image	32	100%
Video	24	75%
Audio	11	34%
GUI	4	12.5%
Embodied	2	6.3%

After deduplicating each model’s modality set, only eight remain:

#	Bundle	Count	Examples
1	Text + image + video	11	GLM-5, SenseNova 6.5 Pro, ERNIE 5.1, Keye-VL-2.0, Hunyuan-Vision, MiniCPM-V 4.5, Pangu 5.5, DeepSeek-V4.1-Flash, Guanlan 3, Fuxi 2, Emu3.5
2	Text + image	6	DeepSeek-OCR 2, BAGEL-2, InternVL 4.0, Cloudwalk Congrong 2, Ant Ling 2, JD Yanxi 2
3	Text + image + video + audio	8	Doubao Seed 2.0 Pro, Qwen3.5-397B, MiMo-V2.5, LongCat-Next, Qwen3-Omni-30B, Step-3, MiniMax M2.5, Zidong Taichu 2.0
4	Text + image + video + GUI	3	Kimi K3, Qwen3.8-Flash, GLM-5.3
5	Text + image + audio	1	Spark X2
6	Text + image + video + audio + embodied	1	WITA-Omni
7	Text + image + video + audio + GUI	1	Qwen3.8-Max
8	Text + image + embodied	1	NeuroXess-N1

Three structural facts:

Text + image is the floor; there are no exceptions. All 32 have both; none is audio-only, video-only or embodied-only. “Multimodal” in this sample means “text + image plus something”, not a replacement for text or image.

Video is the divide. 24 have video, 8 do not; every GUI model has video, every embodied model has image. Video is not “one more modality”; it separates general models from specialists.

Audio and GUI barely meet. 11 have audio, 4 have GUI, and their intersection is one model (Qwen3.8-Max). This is not random: audio comes from speech and call-centre products, GUI from screen-operation products; the data-collection pipelines are completely different.

Negative space one: no 6/6

Under the six-modality taxonomy, no model has all six.

The closest are two 5/6 models:

• Qwen3.8-Max: text + image + video + audio + GUI, missing embodied. An “all-modality model inside the screen”.
• WITA-Omni: text + image + video + audio + embodied, missing GUI. An “all-modality model inside a body”.

One lacks a body, the other lacks a screen — two completely different product routes. Calling both “omnimodal” hides a fact: in current product form, “omnimodal” is at most 5/6, and the missing piece determines which scenario it can serve.

Zidong Taichu 2.0, often described as “all-modality”, is actually 4/6 (text + image + video + audio, missing GUI and embodied); its “all” is the academic four-modality definition, not our six. Different definitions produce different conclusions — and this is where “omnimodal” is most often misread.

Negative space two: embodied + GUI = 0

This is the cleanest zero on the board: no model has both embodied and GUI.

Why it matters: service robots, inspection robots and retail terminals often need two things at once:

1. Understand the physical world: walk to a device, recognise indicator lights, screens, buttons, gauges;
2. Operate a digital interface: tap the screen, enter a password, switch pages, dismiss a dialog.

The first is embodied vision, the second is a GUI agent. The market splits them into two model classes: WITA can see the physical world but cannot tap a screen; Kimi, Qwen3.8-Flash and GLM-5.3 can tap but have no body. The middle step — a robot walks to a screen and completes one touch operation — is covered by no single model.

This is not an argument that vendors should immediately build a 6/6; it is evidence of a real integration gap. A buyer building this scenario needs at least two models plus an orchestration layer, and the budget must include it.

Negative space three: audio + GUI without video = 0

The only model with both audio and GUI, Qwen3.8-Max, also has video. So “audio + GUI, no video” is zero.

It maps to a concrete form: a voice-only screen agent — a call-centre seat listens to the customer while clicking, filling forms and switching pages in the CRM. It does not need camera frames; it needs audio plus screen.

Is this a real gap or a pseudo-gap? We judge it mostly pseudo: Qwen3.8-Max carrying an extra video modality does not prevent audio + GUI work; extra modality is redundancy, not a defect. The real problem is not “the combination does not exist” but “it is not separately evaluated” — audio has no independent dimension, and GUI success rates come from our own tests.

Negative space four: embodied + video without audio = 0

WITA-Omni has video, audio and embodied; NeuroXess-N1 has embodied but no video. So “embodied + video, no audio” is also zero.

It maps to a silent industrial robot: vision only, no microphone, keeping factory noise out of the model. This gap is also mostly pseudo — WITA’s extra audio input can be disabled at deployment. What is actually missing is independent verification that “with audio off, video timing and embodied control still hold”.

Of the four absences, two are real (6/6, embodied + GUI) and two are pseudo (audio + GUI without video, embodied + video without audio). The test is not “does the combination exist” but “does the extra modality block the scenario”. An extra modality usually does not; a missing critical one does.

Modality co-occurrence matrix

	Text	Image	Video	Audio	GUI	Embodied
Text	32	32	24	11	4	2
Image	32	32	24	11	4	2
Video	24	24	24	10	4	1
Audio	11	11	10	11	1	1
GUI	4	4	4	1	4	0
Embodied	2	2	1	1	0	2

The two most glaring zeros: GUI × embodied = 0, and the implied 6/6 = 0. Every other pair has at least one model.

Another number: video × audio = 10, while total audio is 11. Except for Spark X2, every audio model also has video. Audio is almost always a by-product of video in this sample, not an independent capability. That matches the scorecard: video gets 14% weight; audio gets no dimension at all.

Having a modality is not using it

A modality tick is a vendor claim, not proof of capability. Three counterexamples:

Video: 24 have it, but video scores run from 55 to 96. JD Yanxi 2 has video and scores 55; Doubao has video and scores 96. Same tick, a 41-point gap. “Supports video” means the API accepts video, not that it understands timing.

GUI: 4 have it, GUI scores run from 68 to 95. Does BAGEL-2 have GUI? No — its multimodality is text + image. The four GUI models are Kimi, Qwen3.8-Max, Qwen3.8-Flash and GLM-5.3, scoring 95 to 87. GUI is tied for the highest weight (14%) and is one of the least covered modalities: 14% of the weight, 12.5% of the models.

Audio: 11 have it, but there is no independent score. Our eight dimensions do not include audio understanding. This board cannot answer “whose speech is best”; it can only answer “who claims audio support”. That is a gap in our evaluation, and a dimension for the next version.

Embodied: 2 have it, no independent score. WITA and NeuroXess carry embodied capability in scenario descriptions only; their ranks (#23, #30) mostly reflect the other seven dimensions, not embodied ability.

Modality coverage is a necessary condition, not proof of capability. When you see a tick for audio, video or GUI, ask three follow-ups: is there an independent evaluation? What is the accuracy in your scenario? What is the fallback when it fails?

Procurement: use negative space to budget integration

The practical value of modality negative space is not “pressure vendors to build 6/6”; it is calculating integration cost.

Steps:

1. List the modalities your task truly needs. Not “more is better”; which are mandatory.
2. Find coverage among the eight bundles. If none covers it, find the smallest superset.
3. If no superset exists, price two models plus orchestration. Embodied + GUI is the canonical case: one embodied model + one GUI agent + a task layer.
4. Price “extra modality” and “missing modality” separately. Extra modality is usually redundancy, cost is inference; missing modality is a gap, cost is integration.

A concrete example: a service robot needs to check in at a screen. Your set is text + image + video + GUI + embodied. No 6/6 exists, and no minimal superset either (WITA lacks GUI, Qwen3.8-Max lacks embodied). Two routes:

• Route A: WITA sees people and environment; Qwen3.8-Flash (¥1.2, GUI 87) taps the screen; a task orchestrator connects them. Cost: two models plus orchestration.
• Route B: wait for a 6/6. Cost: time, and uncertainty.

Negative space tells you Route A is not a compromise; it is the only feasible engineering answer today. That is why “omnimodal” is attractive in product decks and often means two models in procurement.

Boundaries

• Modality claims come from the modalities field in src/data/models.ts; they are vendor claims, not our measurements. Some models may support more through adapters without declaring it.
• “GUI” and “embodied” are elastic definitions. We define GUI as screenshot understanding plus coordinate actions, and embodied as physical body/sensor input; vendors may use wider definitions.
• 32 Chinese models are not the global sample. The negative space describes this sample only; globally, 6/6 or embodied + GUI may exist.
• Audio and embodied have no independent scores. That is our limitation, not the vendors’. A coverage table cannot replace capability evaluation.

Reproduction

Parse each model’s modalities array from src/data/models.ts, deduplicate to eight bundles; count pairwise intersections for the co-occurrence matrix; derive negative space by subtracting existing sets from the 64 possible subsets of six modalities. Scripts ship with the site data; rerun after every expansion, and the map updates in the changelog.

If this essay leaves one sentence: the multimodal race is not about who ticks the most boxes, but who covers the box nobody else does. And that box is usually not in the launch deck; it is in the integration budget.