Native grades N0–N4: how the 2026 Chinese multimodal tier list reshuffles when only true-native models qualify

The October 2026 edition of the 2026 China native multimodal LLM tier list covers 32 models, and every one of them was called “native multimodal” at its launch event. This is an accounting problem: “native” is the cheapest adjective in the marketing department — it costs nothing to say and makes a parameter table look a generation ahead. Strip the launch language away, look only at architectural evidence, and the 32 models fall into five generations, with three generations between the top and the bottom.

This essay does not re-explain what native multimodality means. It does three things that exist nowhere else: a falsifiable set of criteria from N0 to N4, a placement for all 32 models, and a counterfactual re-ranking that shows what the board looks like under a stricter reading.

Two accounting systems: launch-grade and architecture-grade

The same word means entirely different things under the two systems, and that is where all the arguments come from.

	Launch-grade	Architecture-grade
Definition of “native”	Handles text and images together	Vision and text share one token space and one training objective
Evidence	Launch deck and slides	Tokenizer details, shared decode head, whether the vision tower was trained
Falsifiability	None — “handles” has no threshold	Four questions, each observable
Typical phrasing	“Native multimodal, one model for everything”	“Unified discrete vision tokens; understanding and generation share a head”

Launch-grade flattens three generations of difference into one sentence. Architecture-grade criteria must be falsifiable, or they are just another adjective. We grade with four questions:

1. Are images cut into discrete tokens entering the same vocabulary as text?
2. Do understanding and generation share one decode head?
3. Can gradients flow from the output back into the vision encoder (was the tower actually trained)?
4. Do audio and video share text’s positional encoding and clock alignment?

Grade	Definition	Criteria passed	Typical failure mode
N4	Unified token space across all modalities	All four	Single-modality ceiling limited by discrete compression
N3	Unified discrete vision tokens; understanding and generation share a source	1 and 2	Weaker pixel-level detail than continuous features
N2	Early-fusion continuous features; vision tower jointly trained	Only 3	Loses references in long cross-modal chains
N1	Adapter splicing; vision tower frozen	None of the three; image input only	Subject drift across multi-turn edits
N0	Text model plus an external vision pipeline	Vision lives outside the model	Images are flattened to text before the model sees them

The N2/N1 boundary is not “can it see images” but was the vision tower trained. That is the grade launch language erases most reliably, and it is the most contested placement in this exercise.

Placement results for all 32 models

The spread: two N4, ten N3, fifteen N2, four N1, one N0 — this cycle’s six industry models all land at N2: continuous-feature early fusion is the least-effort architecture choice for an industry model.

Native grade	Model	Overall rank	Basis
N4	Qwen3-Omni-30B	10	Audio, video and text share a token space; official clock alignment disclosed
N4	WITA-Omni Preview	22	Five modalities including embodied; DailyOmni 85.21; audio is not a bolted-on ASR
N3	Qwen3.8-Max	1	Unified vision tokenizer; understanding and generation share a decode head
N3	Qwen3.5-397B	3	Same-generation tokenizer design; 32-language OCR learned at the token layer
N3	GLM-5.3	4	Full Ascend training; vision tower jointly optimised
N3	Kimi K3	5	First model to bring native vision into the main line; CharXiv 84.8
N3	LongCat-Next	9	28× compressed discrete vision; OmniDocBench 0.152
N3	DeepSeek-V4.1-Flash	12	Discrete vision tokens share the text vocabulary; MIT open weights
N3	Qwen3.8-Flash	13	Smaller sibling on Max’s tokenizer generation
N3	Emu3.5	23	Pure next-token line; image, text and video in one space; no audio
N3	BAGEL-2	26	Shared understanding-generation source; image and text only
N3	NeuroXess-N1	25	Neural signals and images in one discrete token space; the only entrant
N2	Doubao Seed 2.0 Pro	2	Continuous vision features plus early fusion; no unified discrete tokenizer disclosed
N2	GLM-5	6	Vision tower trained; generation runs on a separate head
N2	MiMo-V2.5	7	Early fusion; public materials describe no unified output head
N2	SenseNova 6.5 Pro	8	Continuous feature fusion; strong professional vision
N2	ERNIE 5.1	11	Early fusion; video encoded frame by frame
N2	Keye-VL-2.0	14	Continuous feature line with 3B active parameters
N2	Step-3	17	Four-modality interface; architecture details undisclosed
N2	InternVL 4.0	16	ViT early fusion; vision tower jointly fine-tuned
N2	MiniCPM-V 4.5	24	8B edge model; vision tower jointly fine-tuned
N2	Hikvision Guanlan 3	27	Surveillance-video specialist; continuous features + early fusion
N2	Cloudwalk Congrong 2	32	CV lineage; vision-tower led
N2	Zidong Taichu 2.0	25	Academic all-modality; continuous fusion
N2	Ant Ling 2	22	Finance-optimised; continuous fusion
N2	JD Yanxi 2	28	Retail-optimised; continuous fusion
N2	NetEase Fuxi 2	29	Content-optimised; continuous fusion
N1	Hunyuan-Vision	15	Adapter splicing; vision tower frozen (provisional)
N1	Spark X2	18	Speech runs on a separate ASR/TTS pipeline
N1	MiniMax M2.5	19	Interface covers four modalities; architecture undisclosed (provisional)
N1	DeepSeek-OCR 2	20	Layout parsing specialist; vision encoder independent of the language tower
N0	Huawei Pangu 5.5	21	Industry vision is delivered by a separate pipeline at the product layer (provisional)

Four models are marked provisional because their vendors have not published enough architecture. We grade observable behaviour: interleaved image-and-text output in one turn, subject consistency across multi-turn edits, and whether capability collapses with vision removed. The probes are below.

The shadow board: what happens if only N3 and above qualify

This is the core artefact. Exclude N2 and below, and the board shrinks from 32 models to 12:

Shadow rank	Model	Composite	Native grade	Original rank
1	Qwen3.8-Max	95.4	N3	1
2	Qwen3.5-397B	94.6	N3	3
3	GLM-5.3	92.4	N3	4
4	Kimi K3	92.3	N3	5
5	LongCat-Next	89.2	N3	9
6	Qwen3-Omni-30B	88.9	N4	10
7	DeepSeek-V4.1-Flash	87.9	N3	12
8	Qwen3.8-Flash	87.5	N3	13
9	WITA-Omni Preview	80.6	N4	22
10	Emu3.5	80.5	N3	23
11	NeuroXess-N1	77.2	N3	25
12	BAGEL-2	76.9	N3	26

Four conclusions:

• The top spot does not move. Qwen3.8-Max is first under both readings, so its lead does not depend on the “native” premise.
• Doubao Seed 2.0 Pro at #2 is ejected outright. An S-tier model with perception 97 and Chinese 97 has no place under a discrete-token-only reading.
• Tier A+ is emptied. MiMo-V2.5 (#7) and SenseNova 6.5 Pro (#8) are both out, ranks 4 through 8 reshuffle entirely, and the two Flash models move from the teens into the top eight.
• The shadow board’s tail contains the only cold model on it. NeuroXess-N1 enters at N3 — the only model here that puts neural signals in the same token space, and the only one on the shadow board without a public API. Its last two entries (NeuroXess-N1 at 77.2 and BAGEL-2 at 76.9) are both specimens of “architecturally advanced, capability-limited”: native grade preserves their classification, not their ranking.

Why the excluded models scored so high

This is what the shadow board is actually for: not to dismiss Doubao, but to make clear where Doubao’s score comes from.

Doubao Seed 2.0 Pro’s perception 97, Chinese 97 and video 96 all rest on continuous vision features. Continuous features preserve pixel-level fidelity — dish names, crumpled receipts, heritage-site patterns are more stable on them than on discrete token lines. What it loses is the multi-turn edit consistency and long cross-modal chains that come from a shared understanding-generation source.

So the correct reading is not “N2 is worse than N3”:

N2’s score concentrates in perception-type dimensions; N3’s in generation and agent dimensions. The tier list’s weights (perception 20, documents 18, charts/video/agents 14 each, generation 6) happen to favour perception, and that is the direct reason Doubao reaches #2.

Raise the generation weight from 6% to 30% and the N3 camp moves up immediately — we ran that experiment in the 2,000-run weight perturbation, where Emu3.5 alone climbs ten places.

InternVL 4.0 is another N2 specimen worth isolating: perception 96 is the highest open-weights score below the closed flagships, but video is 66. Its architecture (ViT early fusion) spent its entire budget on single-frame understanding — a rational trade in the open-source community, and a 22-place internal spread on this board (perception #4 to video #26).

The cost of native: what discretisation actually loses

“Native is always stronger” is the other slogan. Discretisation compresses images into a finite vocabulary and necessarily loses pixel-level detail. LongCat-Next supplied the useful counterexample in March 2026: with 28× compressed discrete vision it reached 0.152 on OmniDocBench, beating same-period specialist vision models. So discrete must be weaker than continuous is false — what matters is the tokenizer’s semantic completeness, not the compression ratio itself.

But the counterexample covers printed documents only. Where texture matters — hairline scratches in industrial QC, tongue-coating colour transitions, brushwork on artefacts — continuous features still win. That is why neither industrial QC nor tongue diagnosis tops out at an N3 model in the Chinese document tier list nobody benchmarks.

NeuroXess-N1 supplies a third kind of sample: its discretisation target is not natural imagery but EEG/fNIRS temporal signals. Neural signals have no “pixel-level detail” to lose, so discrete tokenisation costs almost nothing — an N3 architecture migrating to a brand-new modality at near-zero cost. Whether native pays always depends on the modality, never on the slogan.

Five probes you can run in ten minutes

You can grade without reading technical reports, at the cost of one notch of precision. Run the five probes in order:

1. Interleaved output probe: ask for “describe this image, then draw an edited version, then explain what changed”. Completing image-text interleaving in one turn is at least N3; needing a separate generation API and stitching is N2 or below.
2. Multi-turn edit consistency probe: give three consecutive edit instructions on the same image. No drift across three turns points to shared state (N3+); changing face on the second points to per-turn re-encoding (N1).
3. Blind-vision collapse probe: replace the image with an equally long text description and compare output quality. A tiny gap means the model was barely using vision — a strong N0/N1 signal.
4. Audio-visual alignment probe (for N4): give a video of someone speaking and ask “who said what, at which second”. Second-level alignment implies audio shares the token space; content-only recap implies a bolted-on ASR.
5. Cross-modal input probe (for cold architectures): feed a non-standard modality’s temporal signal (EEG, EMG, radar returns) and ask for its semantic label. Answering directly puts the modality inside the token space; asking you to convert to text or images first means an external pipeline. Exactly one model on this board passes this probe, and its dossier is here.

None of the probes require vendor cooperation or benchmarks. They cannot separate all of N2 from N3, but they separate N0/N1 from N3/N4 cleanly — which is precisely the cut that matters in procurement.

Why native grade is not in the composite

The most-asked question. The answer is it would double-count.

Architecture’s payoff already shows up in the dimension scores: N3 models score higher on generation, N4 models on video. Adding native grade again scores the same thing twice and tilts the ranking toward a technology line rather than toward capability.

So we publish the grade separately, as a gate rather than a bonus. If your work needs multi-turn image editing or audio-visual alignment, filter out N2 and below; if it is single-turn image understanding only, N2 is the better-value tier.

Both usages are switchable in the model picker. Per-model grade details and placement rationale live on each dossier.

Placements move with evidence

Once the four provisional models publish architecture details, or the community produces reproducible ablations, we will re-grade them and log the change and its reason in the changelog. The native grade is not tenure. Like every score on this board, it is a snapshot under the currently public evidence.