Context Windows in China’s 2026 Model Market: 1M Tokens as Table Stakes and the Stratification Below It

Executive Summary

The 1M-token context window has become table stakes for Chinese frontier models in 2026: ten of the nineteen models in the China AI Hub collection record a 1,048,576-token context window, spanning five of six vendors and both the flagship and flash tiers. Below that ceiling, the market stratifies cleanly — a 256K mid tier (Doubao Turbo, Kimi K2.6/K2.7, Qwen3.8-2.4T-A95B) and a 200K legacy tier (MiniMax M2.7) persist for cheaper or older models. The less-discussed dimension is output: maximum output tokens range from 131K (GLM, Qwen) to 1M (Kimi K3), a 7.6x spread that changes which models can actually generate long documents rather than merely accept long contexts. This article maps both dimensions, explains the tiering logic, and flags the verification gaps.

What We Know

The database records context_window for all nineteen models and maximum_output for ten of them (nine models do not publish an output-token limit — recorded as absent, not as unlimited). The distribution:

1M context (1,048,576 tokens, ten models):

• DeepSeek V4.1-Flash (max output 393,216) and V4-Pro (393,216)
• Doubao Seed 2.1 Pro (262,144) and Seed Evolving (262,144)
• GLM-5.2, GLM-5.3 (131,072 each), GLM-5.3-Flash (131,072)
• Kimi K3 (1,048,576)
• MiniMax M3 (max output not published)
• Qwen3.8-Max (131,072) and Qwen3.8-Flash (131,072)

256K context (four models): Doubao Seed 2.1 Turbo (262,144), Kimi K2.6, Kimi K2.7-Code, K2.7-Code-Highspeed (output limits not published), Qwen3.8-2.4T-A95B.

200K context (two models): MiniMax M2.7 and M2.7-Highspeed.

FACT: 1M context is now the default position for new releases across DeepSeek, ByteDance, Zhipu, Moonshot, MiniMax and Alibaba. Only ByteDance maintains a current-generation model (Turbo) below 1M, deliberately — it is the low-latency tier.

What the Data Shows

Finding 1: context tiering now maps to product tiering, not vendor capability. The 256K/200K models are not failures of capability; they are the cheap or legacy tiers of vendors whose flagships are 1M. MiniMax serves M2.7 at $0.30 alongside M3 at the same price with 1M — the older model’s context is a historical artifact, not a pricing signal. The market has not stratified by “who can do long context” (everyone can) but by “where the vendor draws the product line.”

Finding 2: output tokens are the real differentiator. Among models that publish the limit, the spread is: DeepSeek 393K, Doubao 262K, GLM 131K, Qwen 131K, Kimi K3 1M. A 1M-input model with a 131K output cap (GLM-5.3, Qwen3.8) is not a long-form generation tool; it is a long-input analysis tool. Kimi K3’s 1M input / 1M output pairing is unique in the collection — it is the only model documented as capable of generating a million tokens of output, which reframes its $3/$15 pricing: per-token output pricing on a 1M-output model implies a different cost ceiling for any single request than a 131K-output competitor.

Finding 3: open-weight long context is cheaper but capped. The open Qwen3.8-2.4T-A95B carries 256K — below its closed sibling’s 1M. GLM-5.3-Flash (open, 1M) is the exception that proves the pattern: Zhipu published the open model with the full 1M context, making it the only open-weight 1M-context model in the collection alongside GLM-5.3. Self-hosters chasing long context have a short list.

ANALYSIS: the 1M-input norm and the 131K-1M output spread encode an economic truth: input context is cheap to advertise, output tokens are expensive to serve. Vendors converge on 1M input because the marketing threshold is now there; they diverge on output because output tokens bill at 2-5x input rates and directly consume inference capacity. The output limit is the honest economic parameter.

What Has Changed

The trajectory from the database’s older entries:

• 2025-12 DeepSeek V3.2 and the early 2026 models established the 256K norm. The K2.6 (April 2026) and M2.7 (March 2026) still live there.
• The 1M shift arrived in a six-week burst (June-July 2026): MiniMax M3 (06-01), GLM-5.2 (06-16), Doubao Seed 2.1 family (June/September), Kimi K3 (07-16). By September, Qwen3.8-Max/Flash (August) and DeepSeek V4.1-Flash/V4-Pro completed the sweep.
• Flash models inherited 1M without a price premium. Qwen3.8-Flash ($0.15/$0.47), GLM-5.3-Flash ($0.15/$0.50) and DeepSeek V4.1-Flash ($0.15/$0.60) all carry 1M at the floor price — long context no longer costs extra at the entry tier.
• The open-weight 1M option exists (GLM-5.3, GLM-5.3-Flash) but Alibaba’s open A95B remains at 256K — the open long-context crown currently sits with Zhipu.

Why It Matters

• For long-document analysis (legal, financial, codebase review): any of the ten 1M models works, but output caps decide the deliverable. If the deliverable is a summary, 131K output is ample; if it is a generated report longer than ~100K tokens, only DeepSeek (393K) and Kimi K3 (1M) publish enough headroom.
• For agent workloads: 1M context plus cache pricing is the agent-economics sweet spot — the flash trio (DeepSeek/Qwen/GLM) all combine 1M with cache-hit discounts, which is why agent tooling has standardized on them.
• For self-hosting: the 1M open-weight list is GLM-5.3 (744B/40B) and GLM-5.3-Flash (320B/18B). Nothing else in the collection combines open weights with 1M. This is a real constraint for enterprises that cannot use APIs.
• For latency-sensitive chat: ByteDance Turbo’s 256K at the low-latency tier shows a vendor explicitly trading context for speed — context and latency are still a trade in production, even if the spec sheet says otherwise.

Comparison / Evidence

Normalized view by tier:

Tier	Models	Context	Typical max output
Flagship	V4-Pro, K3, GLM-5.3, M3, Qwen3.8-Max, Seed 2.1 Pro/Evolving	1M	131K-1M (K3 outlier)
Flash	V4.1-Flash, GLM-5.3-Flash, Qwen3.8-Flash	1M	131K-393K
Mid / legacy	Turbo, K2.6, K2.7, A95B, M2.7	200K-256K	mostly unpublished

DATA: compiled from context_window and maximum_output fields, verified 2026-09-20. Nine models do not publish maximum_output; the table records those absences rather than estimating.

Limitations and Uncertainty

Three caveats. (1) Output limits are under-disclosed: nine of nineteen models publish no maximum_output; treating absent as “unlimited” would be wrong, so the analysis distinguishes published caps from unknown ones. (2) Effective context ≠ advertised context: long-context quality degradation is well documented in the field generally, and no vendor in the collection publishes a needle-in-haystack or long-context benchmark score per model — advertised windows are the vendor’s guarantee, not a measured retrieval quality. (3) Version drift: Doubao Seed 2.1 Pro’s 1M context is documented in the September version; earlier snapshots (260628) may differ, and the database records the latest documented value.

China AI Hub View

Our assessment: context length has completed its transition from differentiator to baseline — 1M is now the price of entry for any serious 2026 release, available at $0.15 input. The frontier has moved to two secondary battles: output-token headroom (where Kimi K3’s 1M stands alone) and the economics of long context under load (where cache pricing and linear-attention architectures matter more than the window number itself). For Q4, the database will watch three signals: whether any vendor pushes beyond 1M input; whether open-weight 1M options expand beyond Zhipu; and whether maximum_output disclosure becomes universal — currently the collection’s biggest documentation gap.

Conclusion

The 2026 Chinese market has standardized on 1M-token input across flagships and flash tiers alike, stratified the budget tier at 256K/200K, and quietly left output tokens as the true differentiator — a 131K-to-1M spread that decides what a model can produce, not just what it can read. Kimi K3’s 1M/1M pairing, GLM-5.3-Flash’s open-weight 1M, and the flash trio’s long-context-at-floor-price are the three structural facts of the current window landscape. Buyers should now select on output caps and self-hosting context, because input context no longer discriminates.

Sources

See the Sources list in the page metadata — context and output fields trace to official model pages across all six vendors, verified 2026-09-20. All tiering analysis is flagged ANALYSIS/VIEW and derived only from the recorded fields.