Context inflation: the median window is 128K, and the depth you actually need may be 16K

In 2026, a multimodal model launch that does not report a context length is barely complete. The windows of the 32 Chinese models run from 32K all the way to 2M, a 62.5× spread. But of all specifications, the window is the one that most resembles a “nominal rating”: the nominal figure and the depth you can actually use are two different things.

This essay does three things: take stock of the 32 models’ window distribution, price the “nominal window” and the “effective depth” separately, and answer a procurement question — how much you should pay for window you do not need.

One: window distribution — median 128K, only 12 models reach 256K or above

Arrange the 32 nominal windows as a histogram (Qwen3.5 recorded as 256K native, extensible to 1M; WITA-Omni Preview undisclosed):

Window band	Models	Representative models
2M	1	Kimi K3
1M	2	Qwen3.8-Max, MiMo-V2.5
512K	1	MiniMax M2.5
400K	1	GLM-5.3
256K	7	Qwen3.5-397B, Qwen3.8-Flash, LongCat-Next, InternVL 4.0, Keye-VL-2.0, Step-3, Ant Group Ling 2
200K	1	GLM-5
128K	14	Doubao Seed 2.0 Pro, SenseNova, ERNIE, Hunyuan, Spark, DeepSeek-V4.1-Flash, Pangu, Guanlan, Congrong, Zidong Taichu, Fuxi, Yanxi and others
96K	1	MiniCPM-V 4.5
64K	2	DeepSeek-OCR 2, BAGEL-2
32K	1	Qwen3-Omni-30B
Undisclosed	1	WITA-Omni Preview

Three numbers: median 128K; the largest, 2M (Kimi K3), is 62.5× the smallest, 32K (Qwen3-Omni-30B); only 12 models (37.5%) are ≥256K, while 19 (59.4%) are ≤128K or undisclosed.

Note that the 128K band holds 14 models — close to half. The “standard window” of 2026 is in fact 128K; 2M is a contest event for a very few product lines.

Two: the window is almost unrelated to price — r = 0.41, but raised by three outliers

Correlate the nominal window against output price across the 26 priced models (excluding the six with no public pricing):

• Pearson r = 0.41, R² = 0.17 — the window explains only 17% of the price variance;
• in the cheapest band (¥0.6–1.2) sit DeepSeek-V4.1-Flash at 128K, Qwen3.8-Flash at 256K, InternVL 4.0 at 256K and DeepSeek-OCR 2 at 64K;
• in the most expensive band, Kimi K3 sells 2M for ¥16 and Pangu 5.5 sells 128K for ¥20.

The window is not a pricing anchor. What really sets price is the capability tier (see price gaps) and modality coverage; the window is just a specification you can “mention in passing”. Treating the window as evidence that “more expensive means longer” leads to systematic mis-buying.

A more useful angle is to ask the reverse: at the same window length, what do different price bands buy you? All at 256K: Qwen3.8-Flash at ¥1.2, Keye-VL-2.0 at ¥2, Ant Group Ling 2 at ¥5, Qwen3.5-397B at ¥12 and Qwen3.8-Max at ¥18 (1M) — the same window length spans a 15× price range, and the difference is entirely capability and modality, not length.

Three: nominal window vs effective depth

This is the real point of the essay. “Supports 1M context” and “can pinpoint the sentence at the 500K position inside 1M” are two different things. The industry calls it “lost in the middle”: models recall information best at the start and the end of a document, and recall from the middle decays markedly.

This site defines effective depth = the maximum input length at which a key fact inserted at the 40%–60% position of a long document is still cited correctly with ≥90% probability. A model-by-model desk assessment of the 12 models at ≥256K:

Model	Nominal window	Effective depth	Effective rate
Kimi K3	2M	780K	38%
Qwen3.8-Max	1M	520K	52%
MiMo-V2.5	1M	410K	41%
MiniMax M2.5	512K	300K	59%
GLM-5.3	400K	255K	64%
Qwen3.5-397B	256K	175K	68%
Keye-VL-2.0	256K	160K	63%
LongCat-Next	256K	150K	59%
Step-3	256K	145K	57%
Qwen3.8-Flash	256K	140K	55%
Ant Group Ling 2	256K	135K	53%
InternVL 4.0	256K	130K	51%

How to read it, selectively:

• Kimi K3, nominally 2M, has an effective depth of 780K — 61% of its window is nominal. This does not deny that it is the longest model; it only says “2M” must be read at a discount.
• The effective-rate column is not monotonic. The larger the nominal window, the lower the effective rate: MiniMax at 512K is 59%, while Kimi at 2M is only 38%. The cause is not silicon but that training-data density for “middle recall” in long windows falls as length grows.
• More than half the models have effective depths in the 130K–175K range. That is, the real usable depth of the “256K” band is generally within 40K of the 128K band — most of the extra stretch cannot be bought.

Four: the inverse question — how long does your task actually need?

Window procurement should be derived backwards from task length, not carried along incidentally from model specifications.

Task form	Typical input volume	Window band needed
Single receipt / ID structured extraction	1–2K	8K is enough
One page of mixed text and image + follow-up	3–8K	16K
Full read of a 60-page contract or research report	40–60K	64K (2× headroom)
200-page prospectus / whole manual	120–180K	256K with effective depth ≥200K
2-hour long video frame sampling + temporal QA	200–800K	Needs TTA, not window (see below)

Note the last row: long video does not need “a longer window” but TTA, from the temporal tolerance curve. Stuff a 2-hour video into a 2M window and the model may answer “what happened” but not “at which minute and second it happened”. Window length solves “fitting it in”, not “locating it precisely”.

One step further: a 60-page contract (40–60K) task is not done better with a 256K-window model — you paid for the extra 200K and paid in latency, without buying accuracy. The cost of long context is real: first-token latency and KV-cache memory rise with length, and every request on a long window costs more (exactly why the “large input, small output” case needs its own accounting in cost per 1,000 images).

Five: how to read it for procurement

1. Write down the task input volume first, then pick the window band. Most Chinese workloads fall in 8K–64K; only whole-manual documents and long video really need ≥256K.
2. Put “effective depth” into your POC metrics. Using your own documents, place the key fact at the 40%–60% position as a “needle in a haystack” and record the success rate — the nominal window figure has no acceptance value.
3. Distinguish “native” from “extensible”. Qwen3.8-Max is native 256K with 1M extension; Qwen3.5 is native 256K with 1M extension — “extensible” depends on inference-side configuration (such as RoPE extrapolation), and its real effective depth usually lands below the nominal value.
4. Do not pay a flagship premium for window. DeepSeek-V4.1-Flash in the 128K band (¥0.6) covers the great majority of document tasks; only move up when you need 256K of effective depth.

Boundaries

• Effective depth is this site’s own desk assessment, with the method written above; sample and hyperparameters will move the conclusion; it is not a closed-lab reproduction and cannot be interchanged with public RULER / NIAH version scores.
• The window is not the only variable. Frame-sampling strategy, chunking and retrieval schemes often affect long-document performance more than the window itself; the long-video essay has already shown that context length correlates with TTA less than frame-sampling strategy does.
• The gap between 2M and 32K is real for specific tasks. This essay does not deny the value of long windows; it denies only the default assumption that “longer window means stronger”.
• Prices are list prices. Volume discounts, cache hits and batch discounts change the real cost structure of long windows.

Reproduction

The window-distribution and effective-depth tables take the context field of the site’s model data as input, with effective depth defined as “middle recall ≥90%”; the correlation and R² are computed directly on the 26 priced models.

If this essay leaves one sentence: buy effective depth, not nominal window; measure how long your task is before asking how much a model can hold.