China’s Rapid AI Model Releases Highlight A New Industry Standard: Four Frontier-Class Models In Two Months

Four Chinese AI laboratories released high-end open-weight models in roughly eight weeks, from DeepSeek V4 on April 24 through two mid-June launches, according to a July 13 market report from Thorsten Meyer AI. The rapid sequence matters because it suggests that advanced open models are moving to a weeks-long release cycle while remaining much cheaper to host through an API than leading Western proprietary systems.

The sequence began with DeepSeek V4 Pro and Flash on April 24, followed by MiniMax M3 on June 1. Moonshot AI released Kimi K2.7-Code around June 13, while Z.ai released GLM-5.2 between June 13 and June 16, according to the report. The models are downloadable, and the dispatch said most use MIT or similarly permissive licenses.

DeepSeek V4 was reported as a 1.6-trillion-parameter mixture-of-experts model that activates 49 billion parameters per pass and supports a one-million-token context window. MiniMax M3 combines a long context window with native multimodal functions. Moonshot says Kimi K2.7-Code uses about 30% fewer reasoning tokens than K2.6 during agent tasks, while Z.ai’s 753-billion-parameter GLM-5.2 led the open-weight category on the Artificial Analysis index cited by the dispatch.

On BenchLM’s July composite, DeepSeek V4 Pro scored 87, compared with 93 for the highest-ranked proprietary model. GLM-5.1 scored 83, Kimi K2.6 scored 81 and Alibaba’s Qwen 3.5 397B scored 79. Thorsten Meyer AI described those results as evidence of depth across four Chinese laboratories, while warning that the ranking is a single benchmark snapshot.

Open Models Gain Release Speed

The release pattern gives companies more frequent choices for self-hosted and locally controlled AI. According to Thorsten Meyer AI, hosted access to the Chinese models costs roughly five to 30 times less than Western frontier APIs, depending on the service and workload. Combined with permissive licenses and long context windows, that pricing may make on-premises deployment practical for more organizations.

The shift also changes infrastructure planning. A company selecting hardware or building an internal AI platform can no longer assume that an open-weight model will remain current for a year. A weeks-long update cycle can lower capability costs, but it can also increase the work required for testing, governance and model replacement.

Four Labs Now Share Lead

The Chinese open-weight market is no longer centered on one developer. DeepSeek, Z.ai, Moonshot AI and Alibaba each occupy the upper portion of the rankings cited in the report, with different areas of emphasis: low-cost inference, benchmark performance, long-running agents and broad model availability. Thorsten Meyer AI estimates that four of the five strongest open-weight families now come from Chinese laboratories.

For European organizations, the models present both an option and a dependency. Downloaded weights can run on local infrastructure, but hosted Chinese services process data under Chinese law. Some Western public bodies also restrict Chinese-origin applications or models, while the dispatch said downloadable DeepSeek weights remain legal in the United States despite federal restrictions on the DeepSeek application on government devices.

“The cadence is the signal.”

— Thorsten Meyer AI, July 13 market dispatch

Operating Large Language Models Benchmarking, Deployment, RAG, and Prompt Design (Modern AI Systems Book 5)

AmazonView Latest Price

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

Benchmarks and Licenses May Shift

It is not yet clear whether the reported pace will continue through the rest of 2026. The label frontier-class is also an analytical judgment rather than a shared technical standard, and benchmark scores may not predict performance across specific business workloads, languages or agent systems.

Several commercial and policy questions remain open. The developers could change licenses, API prices or access rules in later releases, and governments may impose new restrictions. Moonshot’s claimed reduction in reasoning-token use also requires independent testing across varied agent tasks.

Independent Tests Will Settle Claims

Developers and model evaluators will now test DeepSeek V4, MiniMax M3, Kimi K2.7-Code and GLM-5.2 across coding, multimodal, reasoning and long-agent workloads. Those results should show whether the benchmark positions translate into lower operating costs and stable production performance.

Organizations considering deployment will also need to track license changes, data-processing rules and government restrictions. The next major Chinese releases will indicate whether the spring 2026 cluster established a lasting weeks-long industry cycle or marked an unusually busy period.

Key Questions

Which four models were released?

The report identifies DeepSeek V4, MiniMax M3, Moonshot AI’s Kimi K2.7-Code and Z.ai’s GLM-5.2.

Are all four models open source?

They are described as open-weight models, meaning their trained weights can be downloaded. That does not always mean the training data, full source code and development process are fully open.

How close are they to proprietary leaders?

BenchLM’s July composite placed DeepSeek V4 Pro at 87, six points behind the proprietary leader’s score of 93. That is one tracker and should not be treated as a universal performance measure.

Why are European companies paying attention?

The models may support lower-cost local deployment and reduce reliance on proprietary APIs. Chinese origin, hosted-data rules and possible public-sector restrictions create a separate dependency risk.

Will the current licenses remain available?

No long-term guarantee has been reported. Current permissive terms apply to the cited releases, but developers or governments could alter future licensing and distribution policies.

Source: Thorsten Meyer AI