TL;DR
Four Chinese laboratories released advanced open-weight AI models between April 24 and mid-June 2026. The pace could widen access to self-hosted AI, but benchmark limits, data rules and future licensing policies create uncertainty.
Four Chinese AI laboratories released four advanced open-weight models between April 24 and mid-June 2026, compressing a major product cycle into roughly eight weeks. The sequence, documented by Thorsten Meyer AI, matters because it points to faster competition on model capability, hosting prices and access to downloadable weights.
DeepSeek V4, offered in Pro and Flash versions, arrived on April 24. According to the source material, DeepSeek described it as a mixture-of-experts model with 1.6 trillion total parameters, 49 billion active parameters per pass and a one-million-token context window under an MIT license.
MiniMax M3 followed on June 1, with a one-million-token context window, native multimodal features and a modified MIT license. Moonshot AI released Kimi K2.7-Code around June 13, presenting it as an agent-focused model that uses about 30% fewer reasoning tokens than K2.6. That efficiency figure is a company comparison rather than an independently established result.
Z.ai’s GLM-5.2 arrived between June 13 and June 16. The supplied analysis describes it as a 753-billion-parameter mixture-of-experts model released under the MIT license. All four releases were downloadable, while hosted access was reported to cost far less than leading Western proprietary APIs.
Four Frontier-Class Open Models in Eight Weeks
China’s Release Cadence Is the Story
Same-day-verified market pulse · July 13, 2026
The production line — spring 2026
The board this week — BenchLM overall score, July 2026
Gift & complication — the European read
The gift
Frontier-adjacent capability, permissive licenses, weeks-long refresh cycle. This cadence is what makes serious on-premises AI economically thinkable in 2026.
The complication
Still a dependency — geopolitical, not technical. Hosted Chinese APIs fall under Chinese data law; many Western agencies won’t touch the weights at all. Licensing generosity is a policy, not a law of nature.
The signal: if your infrastructure strategy assumes open models improve slowly, it’s already wrong. If it assumes the current licensing generosity is permanent, it’s unhedged.
AI model hosting server
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Open Models Narrow the Gap
The release pace suggests that open-weight capability is moving on a weeks-long cycle, rather than through occasional annual launches. For companies deploying models on their own infrastructure, frequent releases and permissive licenses can reduce the cost of adopting local or sovereign AI systems.
BenchLM’s July composite gave DeepSeek V4 Pro a score of 87, six points behind a proprietary leader at 93. The source also cited GLM-5.1 at 83, Kimi K2.6 at 81 and Qwen 3.5 397B at 79. Those results indicate depth across several Chinese laboratories, although they do not directly score every newly released model.
self-hosted AI model download
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Four Labs Build Market Depth
The Chinese open-weight market is no longer centered on one developer. DeepSeek, Z.ai, Moonshot AI and Alibaba now pursue different positions: low-cost inference, benchmark performance, agent efficiency and a broad range of self-hostable models. Thorsten Meyer AI estimates that four of the five leading open-weight families now come from Chinese laboratories.
The supplied analysis says hosted Chinese APIs can be five to 30 times cheaper than Western frontier services. Downloaded weights may support private deployments, but hosted services remain subject to Chinese data law, which can exclude them from regulated workloads in Europe and elsewhere.
“The cadence didn’t — and the cadence is the signal.”
— Thorsten Meyer AI
multimodal AI development kit
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Benchmarks and Policy Risks Persist
It is not yet clear how MiniMax M3, Kimi K2.7-Code and GLM-5.2 compare under identical independent testing. Some cited benchmark scores apply to earlier versions, including GLM-5.1 and Kimi K2.6, while company claims about costs and token savings may depend on workload and configuration.
The durability of permissive licensing is also unknown. Future releases could carry different terms, and governments or enterprises may restrict Chinese-origin models for security or procurement reasons. Open weights reduce reliance on hosted services, but they do not remove supply-chain and policy concerns.
AI model licensing software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
New Tests Will Settle Claims
Independent evaluators are expected to test the new releases across reasoning, coding, multimodal tasks and agent workloads. Buyers will also watch license terms, hosting prices and government policy to determine whether the current release pattern can support long-term deployment plans.
Key Questions
Which four models were released?
The releases were DeepSeek V4, MiniMax M3, Moonshot AI’s Kimi K2.7-Code and Z.ai’s GLM-5.2.
Are the models open source?
They are described as open-weight models, meaning their trained weights can be downloaded. That is not always equivalent to full open source, which may also require public training data, code and detailed methodology.
Do benchmarks show they match closed models?
Not conclusively. DeepSeek V4 Pro scored 87 on BenchLM’s July composite, compared with 93 for the proprietary leader, but one benchmark cannot establish performance across every use case.
Can European companies deploy these models locally?
Downloadable weights can support on-premises deployment, subject to licensing, security reviews and applicable law. Organizations handling sensitive data should consult qualified legal, security and compliance professionals before deployment.
Why is the release pace important?
A faster cadence can bring lower prices and better capabilities to self-hosted systems more quickly. It can also shorten evaluation cycles and increase dependence on licensing policies that may change.
Source: Thorsten Meyer AI
