Temporal tolerance curves: turning video understanding into a curve instead of one number on the tier list

The video dimension on the 2026 China native multimodal LLM tier list is a single 0–100 score. This essay explains why that score hides the most important information in video understanding, and how we plan to supply it.

The replacement is a new metric: the temporal tolerance curve, and its area, TTA.

What a single score hides

Two models at video 96 and 94 look two points apart in your eyes. But real video QA never asks for “the right answer” — it asks for “the right answer with the right timestamp”.

A concrete example: a 30-minute meeting recording, asked “who raised the budget objection, and when”.

• Tolerate ±60 seconds and most models above video 85 answer correctly
• Tolerate ±10 seconds and half as many do
• Demand ±1 second and only 4 of 32 do it reliably

The same model might score 92%, 61% and 18% across those three bands. One number flattens three completely different capabilities into an average.

There is a deeper issue: video capability is actually sold as three product forms, and a single score conflates them:

Product form	Deliverable	Time precision	Price tolerance
Summary product	Content tags, rough-cut suggestions, moderation verdicts	±60s	Priced like the Flash tier
Localisation product	Chapter cuts, highlight reels, meeting minutes	±10s	Priced mid-range
Timing product	Subtitles, sports editing, forensics, ad insertion	±1s	Priced like the flagship tier

Buying the wrong form is structurally expensive: using a summary product for timing work does not degrade accuracy a little — it drops from 90% to 18%. So step one of selection is not comparing scores, it is admitting which product you are buying.

The definition of TTA

The temporal tolerance curve plots required time precision (±60s to ±1s, five bands) on the x-axis against answer correctness on the y-axis. The area under the curve is TTA (Temporal Tolerance Area), normalised to 0–100.

• A flat curve: the model really localises events; accuracy decays slowly as precision tightens
• A steep curve: the model is only doing coarse summaries and collapses the moment exact timestamps are required

TTA’s value is that it separates the two capabilities the video score blends together.

So long-video specialists are not penalised, we compute by duration band and then merge: short clips (< 3 min, weight 30%), 30-minute (weight 45%) and 2-hour (weight 25%). Each band gets its own curve and TTA is the weighted area across the three.

Measured decay: ±60s to ±1s

Event-localisation correctness in the 30-minute band (desk assessment from vendor disclosures, community reproductions and sampling):

Model	±60s	±10s	±3s	±1s	Curve shape
Doubao Seed 2.0 Pro	97	91	82	68	Flat
Keye-VL-2.0	95	89	79	61	Flat
Qwen3.8-Max	96	86	71	52	Medium
MiMo-V2.5	94	82	63	41	Medium
GLM-5.3	92	80	62	44	Medium
Qwen3.5-397B	93	79	60	39	Medium
WITA-Omni Preview	91	78	66	55	Flat
Kimi K3	90	70	45	24	Steep
Qwen3-Omni-30B	89	68	44	26	Steep
DeepSeek-V4.1-Flash	82	51	26	11	Extremely steep
Qwen3.8-Flash	84	55	30	13	Extremely steep
NeuroXess-N1	74	40	14	4	Extremely steep
InternVL 4.0	76	42	16	5	Extremely steep
Hikvision Guanlan 3	92	66	40	18	Steep

The last two columns are the real point of this table. Kimi K3 scores 90 at ±60s, looking close to MiMo-V2.5’s 94; at ±1s it is 24 against 41. Reading only the video dimension scores (Kimi 90, MiMo 93), you cannot see that gap at all.

WITA-Omni Preview is the counterexample: 22nd overall, 6th on video, yet one of the two flattest curves here — 55 still standing at ±1s. The reason is its N4 native grade: audio-visual alignment is learned by the model rather than stitched by an external timestamp pipeline. For subtitles, meeting minutes and sports editing it is far more usable than its overall rank suggests.

This cycle’s entries split into three video bands: NeuroXess-N1 (70) and InternVL 4.0 (66) barely clear the ±60s band — frame-level description only, a product-positioning fact rather than a defect (one does neural signals, one does open-source perception, and neither budgeted for long-horizon timing); Fuxi 2 (84) and Zidong Taichu 2.0 (84) sit on the ±10s threshold — content-oriented timing works, precise timing does not; Guanlan 3 (91) is the only industry video entry, with the curve shape above.

TTA re-ranking: up to 6 places off the video dimension

Re-sorting by TTA against the tier list’s video-dimension ranks:

Model	Video score	Video rank	TTA	TTA rank	Change
Doubao Seed 2.0 Pro	96	1	90.4	1	0
WITA-Omni Preview	92	6	85.8	2	+4
Keye-VL-2.0	94	2	85.1	3	−1
MiMo-V2.5	93	3	76.2	4	−1
GLM-5.3	88	11	75.5	5	+6
Qwen3.8-Max	93	4	74.9	6	−2
Qwen3.5-397B	92	5	71.8	7	−2
Kimi K3	90	7	58.6	8	−1
Hikvision Guanlan 3	91	7	54.0	9	−2
DeepSeek-V4.1-Flash	78	17	38.2	15	+2

Guanlan deserves its own paragraph: its video score of 91 ranks 7th — a single point behind WITA-Omni’s 92 — yet its TTA is only 54.0, ranking 9th: flat on coarse events, steep at second precision. That is not a defect; it is the product form of surveillance video. Perimeter-intrusion review needs minute-level recall, and second precision is a bonus. It manages only 18% in the ±1s band, against Doubao’s 68% at a similar 90-point video score. Two models with similar video scores can differ 3.8× in usability as a timing product — which is the entire reason TTA exists.

Two risers deserve their own paragraphs:

WITA-Omni Preview gains four places. One of the flattest curves on the site, because it is N4 — audio, video and text in one token space, with time alignment learned rather than assembled. Its low overall rank (22) comes from ecosystem and cost: no public API pricing, aimed at embodied interaction.

GLM-5.3 gains six places. Its video score is only 88 (level with GLM-5), but the curve is far flatter. That is the same fact as its agent score of 95 seen from the other side: GLM-5.3 maintains mid-task state well, and that capability transfers directly to long-video temporal consistency.

Context length is not the answer

A common misconception: longer context, better video understanding. The data says no.

Kimi K3’s context is 2M, the longest here; its TTA is 58.6, 8th. Qwen3-Omni-30B has only 32K and scores slightly higher. The reason is that stuffing more frames into context dilutes attention: a 30-minute video at 1 fps is 1,800 frames, each hundreds of tokens, and the model must find one timestamp among hundreds of thousands of vision tokens while positional-encoding extrapolation error accumulates with length.

What actually matters is frame sampling and hierarchical summarisation:

• Coarse sampling to locate the candidate interval → fine sampling to pinpoint → dense encoding only for the candidate interval
• Or convert the video to a timestamped event stream first, then reason over the event stream

Either approach lets a 32K-context model beat a 1M-context model at the ±3s band. Context length decides how much you can stuff in; sampling strategy decides whether what you stuffed in is useful.

For converting each model’s context into usable minutes under one sampling strategy, see the context converter — comparing minutes is more meaningful than comparing K.

Choose by your tolerance

Do not ask “which video model is strongest”; ask “how precise a timestamp do I need”.

Your time tolerance	Usable models	Recommendation
±60s (summaries, rough cuts)	19 of 32	Take the cheapest: Qwen3.8-Flash, DeepSeek-V4.1-Flash
±10s (chapter cuts, highlight reels)	12	MiMo-V2.5, GLM-5.3, Qwen3.5-397B
±3s (subtitles, meeting minutes)	7	WITA-Omni, Doubao, Keye-VL-2.0
±1s (sports editing, forensics, ad insertion)	4	Doubao Seed 2.0 Pro, WITA-Omni, Keye-VL-2.0, GLM-5.3

The last row is the most expensive tier on this board: demand second-level alignment and every cheap Flash option is out, leaving four models — one of which (WITA-Omni) has no public API pricing.

If your business really needs ±1s, budget for the flagship tier, not the average tier. That is the most common budgeting error in video selection. Conversely, if your business is moderation and rough cuts, price for the Flash tier — 19 models are usable at ±60s and the cheapest costs ¥0.54 per thousand images. Paying flagship prices for precision you will never use is the second most common error.

The full TTA computation table and band weights ship with the site data. It is a desk assessment, not a closed-lab reproduction — the same boundary as every score here, written up on the methodology page.