China’s AI Revolution: Four Frontier-Class Open Models Unveiled In Only Eight Weeks

Four Chinese AI laboratories released advanced open-weight models in roughly eight weeks between April 24 and mid-June 2026, according to a July 13 review by Thorsten Meyer AI. The releases from DeepSeek, MiniMax, Moonshot AI and Z.ai matter because their downloadable weights, long context windows and lower hosted prices could make high-capability AI more accessible outside closed Western platforms.

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 introduced GLM-5.2 during the same mid-June period. Thorsten Meyer AI describes all four as downloadable and says most carry MIT or modified-MIT licensing, though organizations must review the terms attached to each release.

DeepSeek V4 is described as a 1.6-trillion-parameter mixture-of-experts model activating 49 billion parameters per pass, with a one-million-token context window. MiniMax M3 combines a similarly long context window with native multimodal functions. Moonshot positions Kimi K2.7-Code for agent-based coding work and says it uses about 30% fewer reasoning tokens than K2.6. Z.ai’s GLM-5.2 is reported as a 753-billion-parameter mixture-of-experts model.

The review says hosted access to the Chinese models costs five to 30 times less than Western frontier APIs, depending on the service and workload. That comparison is a market estimate rather than a universal price ratio: token categories, caching, hosting arrangements and usage volumes can produce different costs.

Open Models Gain Deployment Weight

The release schedule suggests that high-capability open-weight models are being refreshed over weeks rather than annual product cycles. For companies operating private infrastructure, the combination of downloadable weights, permissive licenses and long context windows can reduce dependence on a single hosted provider and make local deployment more economically feasible.

The competitive depth may matter as much as the pace. BenchLM’s July composite placed DeepSeek V4 Pro at 87, six points behind a closed model scoring 93. It listed earlier models from Z.ai, Moonshot AI and Alibaba close behind. Thorsten Meyer AI concludes that four of the five leading open-weight families now come from Chinese laboratories, but that finding depends on the chosen benchmark and its weighting.

European organizations face a mixed calculation. Self-hosted weights can support local processing and infrastructure control, while hosted Chinese services may expose prompts to Chinese data-law requirements. Model origin, security review and procurement policy may also prevent some government agencies or regulated companies from adopting the software even when local hosting is technically possible.

China Builds Open-Model Depth

China’s open-weight market previously centered heavily on DeepSeek and Alibaba’s Qwen family. The latest sequence adds competitive models from MiniMax, Moonshot AI and Z.ai, with each laboratory emphasizing a different area: low inference prices, multimodal processing, long-running agents or benchmark performance.

Alibaba’s Qwen models remain relevant even though they were not part of the four-release sequence. The family includes smaller variants suited to local hardware and models distributed under the Apache 2.0 license. The source contrasts this growing Chinese field with a thinner Western open-weight market, while identifying Ai2’s Olmo line as a more fully open alternative that releases additional training materials.

The report links some of China’s efficiency work to restricted access to advanced US chips, but that explanation remains an interpretation rather than a confirmed cause for each model design. Commercial competition for developers and infrastructure customers offers another explanation for the rapid schedule and low API prices.

“The cadence didn’t — and the cadence is the signal.”

— Thorsten Meyer AI, July 13 market review

Rankings and Licenses Remain Fluid

It is not yet clear whether all four models will reproduce their reported performance across independent tests, languages and production workloads. BenchLM’s figures represent a single July composite snapshot, while the claim that GLM-5.2 leads open-weight models comes from the Artificial Analysis index. Different evaluations can rank the same models differently.

The durability of current access is also unknown. Permissive licensing is a provider policy that can change in later releases, and open weights do not automatically disclose training data or provide a complete development record. The source does not establish whether API discounts will persist after providers gain market share or face higher operating costs.

Release details also require model-by-model checks. The supplied review gives a date range of June 13 to June 16 for GLM-5.2, rather than one definitive launch date, and does not specify the exact license attached to Kimi K2.7-Code. Those points should be confirmed through each laboratory’s repository and service documentation.

Independent Tests and Enterprise Reviews

Developers are likely to compare the releases through independent coding, agent and multimodal evaluations, while enterprises test latency, hardware requirements and total operating cost. Security teams will also examine model provenance, software dependencies and data handling before approving production use.

The next evidence will come from whether Chinese laboratories maintain this weeks-long release pace, whether licenses remain permissive and whether Western developers answer with stronger open-weight systems. Procurement decisions will depend less on headline benchmark scores than on repeatable performance, legal review and deployment cost.

Key Questions

Which four models were released?

The sequence covered DeepSeek V4 Pro and Flash, MiniMax M3, Moonshot AI’s Kimi K2.7-Code and Z.ai’s GLM-5.2. Their reported release dates ran from April 24 through mid-June 2026.

Are these models fully open source?

They are described as open-weight models, meaning users can download model weights. That does not always include training data, training code or full development records, so open-weight and fully open-source are not interchangeable labels.

How close are they to closed frontier models?

On BenchLM’s July composite, DeepSeek V4 Pro scored 87 against 93 for the listed proprietary leader. That six-point gap applies only to that tracker and should not be treated as a universal measure of real-world performance.

Can European companies use the models locally?

Downloadable weights may allow self-hosting on private infrastructure, subject to license, hardware and security requirements. Hosted Chinese APIs create a different risk profile because prompts may be processed under Chinese law.

Why is the eight-week schedule important?

It indicates that advanced open-weight capability is improving rapidly across several competing laboratories. If sustained, the pace could shorten purchasing cycles and push organizations to build flexible, model-independent infrastructure.

Source: Thorsten Meyer AI