The Breakthrough In AI: China Launches Four Frontier-Class Open Models In Just Two Months

Chinese AI laboratories released four high-end open-weight models in roughly eight weeks, from DeepSeek V4 on April 24 to new systems from MiniMax, Moonshot AI and Z.ai in June, according to a July 13 report from Thorsten Meyer AI. The compressed schedule matters because it suggests that downloadable models are improving on a weeks-long cycle, giving businesses more alternatives to costly proprietary services.

The reported sequence began with DeepSeek V4 Pro and Flash, 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. All four systems were described as downloadable, with most offered under MIT or modified-MIT licenses.

DeepSeek V4 was reported to use a mixture-of-experts architecture containing 1.6 trillion total parameters while activating 49 billion for each pass. DeepSeek and MiniMax also advertised context windows of up to one million tokens. The report described MiniMax M3 as natively multimodal and Kimi K2.7-Code as a system designed for extended coding-agent tasks.

Thorsten Meyer AI said hosted access to the Chinese models costs five to 30 times less than leading Western proprietary APIs, although the report did not provide a normalized price table covering identical workloads. It also cited a July BenchLM score of 87 for DeepSeek V4 Pro, compared with 93 for an unnamed proprietary leader. Those figures represent one benchmark provider’s composite, not a settled measure of general capability.

Open Models Accelerate Deployment Choices

The release pace could change purchasing decisions for organizations that run AI on their own infrastructure. Downloadable weights and permissive licenses can reduce reliance on a single API provider, while lower inference prices may make large-scale local deployment viable for more companies, research groups and public institutions.

The report also points to a broader change in the open-weight market. It says four of the five strongest model families now come from Chinese laboratories: DeepSeek, Z.ai, Moonshot AI and Alibaba. If the benchmark comparisons hold across independent tests, competition would no longer depend on one standout Chinese developer but on several laboratories pursuing different technical strategies.

China Builds a Deeper Model Bench

Chinese developers have expanded their open-weight offerings as access to the most advanced AI chips remains constrained by United States export controls. The source interprets some recent efficiency work as a response to hardware scarcity, though it also describes the releases as an effort to make Chinese systems a default foundation for global AI applications.

The four laboratories emphasize different uses. DeepSeek competes heavily on price and sparse-model efficiency; Z.ai targets benchmark performance; Moonshot AI focuses on long-running agents; and Alibaba’s Qwen range includes smaller models suited to local hardware. This variety gives users choices beyond a single large general-purpose model.

“That’s not a wave. That’s a production line.”

— Thorsten Meyer AI, in its July 13 AI Dispatch

Benchmarks and Licensing Need Scrutiny

Several claims still require independent testing. The label frontier-class is an assessment, not a standardized designation, and benchmark rankings may change with different prompts, evaluators and task mixes. The source excerpt also compares the newly released GLM-5.2 with a BenchLM score for GLM-5.1, leaving GLM-5.2’s placement on that table unclear.

Licensing and regulatory exposure also vary by deployment method. Locally hosted weights do not send prompts to a provider by themselves, but organizations must still review model provenance, security controls and license terms. Using a hosted Chinese API creates a different risk profile because prompts may be processed under Chinese law. It is also unknown whether future releases will retain the same licensing terms.

Independent Tests Will Settle Rankings

Developers and enterprise buyers will now watch for independent evaluations of the four releases, including coding reliability, agent stability, multimodal performance, hardware requirements and total operating cost. Updated BenchLM and Artificial Analysis results should clarify where GLM-5.2 and Kimi K2.7-Code rank, while future release terms will show whether the current open-weight cadence and licensing policies persist.

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 these models fully open source?

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

Do the models match leading proprietary AI systems?

Not conclusively. The cited BenchLM table placed DeepSeek V4 Pro six points behind its proprietary leader, but one composite benchmark cannot establish parity across every task.

Can European organizations use these models locally?

Local deployment may be technically possible under the applicable license, but organizations must examine privacy, security and procurement rules. Hosted APIs require separate review because data leaves the organization’s own environment.

Why is the eight-week schedule important?

It indicates that high-end open models may now refresh every few weeks. That faster cycle can lower costs and expand choice, while forcing buyers to reevaluate models more frequently.

Source: Thorsten Meyer AI