How DeepSeek Changed China's AI Market: Open Weights, Price Floors and the Standard It Set

Every factual claim in this article is drawn from the China AI Hub database (models, pricing and company records, last verified 2026-09-20) and the primary DeepSeek sources linked at the bottom. Benchmark figures are vendor-reported and labeled as such. Where the database records a field as null or undisclosed, the article says so rather than inferring.

Executive Summary

DeepSeek’s influence on China’s AI market is not primarily about any single model release. It is about resetting the baseline that its five rival frontier labs now have to answer to. The China AI Hub data shows DeepSeek converging the market on three standards at once: permissive MIT open weights with no revenue thresholds; a sub-$0.20 flash-tier API price; and a 1M-token context window with 384K maximum output. The clearest evidence is comparative — Qwen3.8-Flash and GLM-5.3-Flash both list a $0.15 input price, exactly matching DeepSeek’s V4.1-Flash off-peak rate. DeepSeek also introduced the only peak/off-peak pricing scheme in the six-provider database, and its September 2026 reversal of the V4-Pro retirement showed how much its API customers push back when continuity is threatened.

What We Know

DeepSeek’s public model history in the database runs from DeepSeek-V3.2 (2025-12-01, MIT) through the V4 family. The V4 Preview opened weights on 2026-04-24 with a 1.6T-total / 49B-active mixture-of-experts design. The GA checkpoint (V4-Pro-0813) shipped on 2026-08-13. On 2026-09-10 DeepSeek released V4.1-Flash — a 552B-parameter MoE that activates only 8B parameters per token on input (16B on output) — and announced MIT open weights for it, plus a reversal of the V4-Pro API retirement.

Three facts anchor the story:

• Every DeepSeek model in the database is MIT-licensed. V3.2, V4-Pro and V4.1-Flash all carry MIT open weights with no revenue trigger, no attribution mandate and no Model-as-a-Service carve-out. This is the permissive end of the Chinese licensing spectrum — the same models that, at Zhipu, MiniMax and Moonshot, carry conditional terms (Apache-2.0 with per-card verification notes, or custom licenses with revenue thresholds) are simply MIT at DeepSeek.
• DeepSeek publishes the market’s cheapest flash tier. V4.1-Flash lists $0.15 input / $0.60 output per 1M tokens off-peak, with a $0.003 cache-hit rate. Its flagship V4-Pro lists $0.66 / $1.98 off-peak.
• Peak/off-peak pricing is a DeepSeek innovation in this dataset. Introduced effective 2026-08-16, peak hours (01:00–04:00 and 06:00–10:00 UTC weekdays, excluding Chinese public holidays) are billed at 2x off-peak. No other provider in the six-provider database uses this structure.

What the Data Shows

The price anchor is the most measurable effect. Comparing input prices per 1M tokens across the database’s flash/entry tiers:

Provider	Flash-tier model	Input / 1M	Output / 1M
DeepSeek	V4.1-Flash	$0.15	$0.60
Alibaba	Qwen3.8-Flash	$0.15	$0.47
Zhipu	GLM-5.3-Flash	$0.15	$0.50
MiniMax	M3 (standard tier)	$0.30	$1.20
Moonshot	Kimi K2.7-Code	$0.95	$4.00

The $0.15 input figure appears three times, at three different companies. That is not coincidence; it is a converged floor. DeepSeek’s V4.1-Flash and its two rivals’ flash models all list $0.15 input. (Output rates differ — Alibaba $0.47 and Zhipu $0.50 undercut DeepSeek’s $0.60 — so the convergence is on the input side, which is what drives batch ingestion, RAG and long-context workloads.)

Context length shows the same pattern. DeepSeek’s V4 family standardized on a 1M-token context and 384K maximum output. In the 19-model database, 11 models now offer a 1M context window, and the V4-Pro-era output ceiling of 384K exceeds every other provider except Kimi K3, which matches it at 1M input / 1M output.

What Has Changed

Three events in the database timeline mark DeepSeek’s trajectory.

First, the V3.2-Exp release (2025-09-29) introduced DeepSeek Sparse Attention and cut API prices by more than 50% — the earliest recorded instance in this dataset of DeepSeek using a price cut as a competitive instrument. Second, the V4 Preview (2026-04-24) put 1.6T-parameter open weights into the market under MIT, which is when the open-vs-closed split in China’s frontier tier became impossible to ignore. Third, the V4.1-Flash release (2026-09-10) shipped a model the vendor describes as using an “asymmetric architecture” — 552B parameters but only 8B active on input — with a claim that KV cache needs one-quarter the HBM and one-eighth the SSD storage of the previous generation.

The most revealing event is the V4-Pro retirement reversal. On 2026-09-10 the change log reversed an earlier statement that V4-Pro requests would route to V4.1-Flash after 2026-09-14. The updated text says V4-Pro API service continues “in response to user demand,” with billing unchanged. A company that controls its own pricing does not reverse a deprecation mid-flight unless the installed base objects loudly. That reversal is evidence of real API customer lock-in, not just marketing.

Why It Matters

For anyone evaluating the Chinese model market, DeepSeek’s strategy forces a specific reading. MIT licensing means DeepSeek’s weights can be rehosted, fine-tuned and monetized by competitors without a revenue conversation. That puts downward pressure on the price of capability itself — a rival cannot charge a premium for a capability DeepSeek has already given away under MIT unless it offers something the weights alone do not provide (closed multimodal features, service guarantees, or tool ecosystems).

The practical consequence is that DeepSeek functions as the market’s reference implementation. The database’s own comparison records and pricing pages show other labs anchoring their flash-tier pricing to DeepSeek’s, and their licensing terms to the question DeepSeek’s MIT license raises: if the cheapest frontier-adjacent model is MIT, why would a buyer accept a revenue-threshold license for a model that is not clearly cheaper or more capable?

Detailed Analysis

We see DeepSeek’s influence as operating on four distinct axes.

Licensing. DeepSeek is the only one of the six labs whose entire open-weight catalog is MIT with no conditions. Zhipu’s GLM models are Apache-2.0 but the database notes the GitHub repo metadata is the license source and the README carries no separate weights-license section — a caveat a commercial adopter must verify per-model. Moonshot’s Kimi K3 uses a custom license that requires a separate agreement for Model-as-a-Service operators above $20M in 12-month revenue, and UI attribution above 100M MAU or $20M monthly revenue. MiniMax’s M-series uses non-commercial or attribution-plus-authorization terms. DeepSeek asks for none of this.

Pricing mechanics. Peak/off-peak is the structural outlier. It is a demand-shaping tool: by making peak capacity 2x the price, DeepSeek pushes batch and non-urgent workloads into off-peak hours and reserves peak capacity for latency-sensitive traffic. No rival in the database has copied it — which is itself notable, because the database shows rivals did copy the $0.15 flash floor.

Architecture. The V4.1-Flash “asymmetric” design — huge total parameter count, tiny active count — is a cost engineering decision made visible. An 8B-active model with a 552B MoE backbone can advertise frontier-scale capacity while serving tokens at a fraction of the compute cost. If the vendor’s KV-cache claims (1/4 HBM, 1/8 SSD) hold, the unit economics of serving long context drop sharply, which is what lets the $0.15/$0.60 price exist at all.

Open infrastructure. DeepSeek’s GitHub organization ships training and inference kernels — FlashMLA, DeepGEMM, DeepEP, DualPipe, 3FS — under MIT, alongside the models. This matters less to end buyers than to the ecosystem: it lowers the cost floor for anyone who wants to serve DeepSeek-style models themselves, which in turn disciplines the hosted API price.

Comparison / Evidence

The strongest comparative evidence is the licensing table. Across the open-weight models in the database:

Model	License	Commercial trigger
DeepSeek-V4-Pro / V4.1-Flash / V3.2	MIT	None
GLM-5.3 / 5.3-Flash	Apache-2.0	None (but per-card verification flagged)
Kimi K3	Kimi K3 License	>$20M/12mo MaaS; >100M MAU or >$20M/mo display
MiniMax M3	MiniMax Community	>$20M/yr revenue → written authorization
MiniMax M2.7	Custom non-commercial	Any commercial use → prior authorization
Qwen3.8-2.4T-A95B	Qwen3.8-Max License	>100M MAU or >$20M/mo display; >$50M/12mo MaaS

The pattern is clear: DeepSeek is the only lab whose open models carry a license that a Fortune-500 legal team can clear in one pass. That is a genuine moat in the enterprise segment, and it is a moat built on not extracting revenue from licensing.

Benchmark evidence must be treated carefully. DeepSeek reports V4.1-Flash at GPQA Diamond 90.9, Terminal-Bench 2.1 90.6 and DeepSWE v1.1 74.2, and V4-Pro at HLE 42.7 (60.0 with tools) and Terminal-Bench 2.1 87.9 — all vendor-reported, with the caveat that DeepSeek’s code-agent benchmarks used DeepSeek Harness in minimal mode and DSBench-FullStack/Hard are internal test sets. These numbers are not independently verified, and they should not be compared across vendors without a shared methodology (the database’s own benchmark-methodology research covers this problem).

Limitations and Uncertainty

Several points constrain this analysis. DeepSeek’s serving regions are not stated in official documentation, so any claim about where inference runs is unverifiable from primary sources. The Hugging Face weights for V4-Pro were last modified 2026-06-22, and the database flags that it is unclear whether they match the 0813 GA checkpoint — a material uncertainty for anyone self-hosting. The vendor benchmark figures are not independently reproduced. And the database’s company record does not establish a founding date or verbatim headquarters line from official pages, so historical framing about DeepSeek’s origin is deliberately kept to what the primary sources state. Finally, the “changed the market” framing is our interpretation of convergent pricing and licensing, not a controlled causal claim — multiple labs could have converged on $0.15 flash pricing independently.

China AI Hub View

Our assessment: DeepSeek’s durable contribution to China’s AI market is the permissionless floor — MIT weights, a $0.15 flash price and a 1M context default that force every rival to justify charging more or gating more. The peak/off-peak scheme is the one piece of DeepSeek’s pricing the market has not copied, and we read that as the difference between a price cut (easy to match) and a pricing mechanism (harder to match without the same demand profile). We expect the V4-Pro continuation to be temporary — the deprecation signal itself remains in the database — but the fact that DeepSeek felt compelled to reverse it is a measure of how dependent its API customers have become. The next test of DeepSeek’s influence is whether the V4.1-Pro launch (referenced in the change log as the eventual V4-Pro replacement) holds the MIT-and-$0.15 line or finally introduces conditions of its own.

Conclusion

DeepSeek changed China’s AI market by making the cheapest credible option also the most permissively licensed one. The evidence is in the database itself: a $0.15 flash price that two rivals matched, an MIT license regime no rival fully matches, a 1M/384K context standard that spread across 11 of 19 models, and a pricing mechanism — peak/off-peak — that remains uniquely DeepSeek’s. For buyers, the lesson is that “Chinese open model” is not one thing; DeepSeek’s MIT terms are the reference point against which every conditional license should be measured, and every benchmark claim from any vendor, DeepSeek included, should be read as vendor-reported until independently verified.