TL;DR
Chinese labs released four high-end open-weight AI models between April 24 and mid-June 2026, according to Thorsten Meyer AI. The pace could lower self-hosting costs, but benchmark limits, regulatory barriers and future licensing policies leave open questions.
Chinese AI laboratories released four high-end open-weight models between April 24 and mid-June 2026, according to a July report from Thorsten Meyer AI, compressing what had often been an annual upgrade cycle into roughly eight weeks. The releases from DeepSeek, MiniMax, Moonshot AI and Z.ai could reshape the cost and availability of capable systems for companies seeking alternatives to proprietary Western services.
The sequence began with DeepSeek V4 on April 24, followed by MiniMax M3 on June 1. Moonshot AI’s Kimi K2.7-Code and Z.ai’s GLM-5.2 arrived within days of each other in mid-June, the report said. All four were described as downloadable, while most carried MIT or modified-MIT licensing terms.
Thorsten Meyer AI said DeepSeek V4 uses a mixture-of-experts architecture with 1.6 trillion total parameters and 49 billion active during each pass, alongside a one-million-token context window. The report described MiniMax M3 as a low-cost, multimodal model with the same context length, Kimi K2.7-Code as an agent-oriented coding model and GLM-5.2 as a 753-billion-parameter mixture-of-experts system.
In BenchLM’s July composite, as cited by the report, DeepSeek V4 Pro scored 87, six points behind a proprietary leader at 93. GLM-5.1 scored 83, Kimi K2.6 reached 81 and Qwen 3.5 397B scored 79. Those results are a single benchmark snapshot, and some ranked versions differ from the models in the release sequence.
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.
Personal AI Servers: A Guide to Building Private AI Infrastructure for Secure, Offline and Self-Hosted Local LLMs for Data Privacy
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Open Models Enter a Faster Cycle
The pace matters because organizations can now receive major open-weight upgrades within weeks, rather than planning around infrequent model generations. Faster releases may reduce the capability gap between locally deployed systems and closed services while putting downward pressure on hosted API prices.
For European companies pursuing local or sovereign deployments, downloadable weights and permissive licenses can make on-premises AI more economically practical. Thorsten Meyer AI estimated that hosted Chinese services were five to 30 times cheaper than Western frontier APIs, although the supplied material did not provide a standardized price comparison or workload assumptions.
The development also reflects depth across several suppliers. The report identified DeepSeek, Z.ai, Moonshot AI and Alibaba as upper-tier open-model developers with different focuses, including pricing, agent reliability and smaller self-hosted variants. That diversity may reduce reliance on any single laboratory, but it still leaves users dependent on Chinese-origin model ecosystems.
Accelerate Everything with Tensor Cores: A Developer’s Guide to High-Performance AI, Efficient Training, and Scalable Models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
China Broadens Its Open-Model Lead
Chinese laboratories have increasingly used open-weight distribution to expand adoption beyond their domestic market. Thorsten Meyer AI said four of the five strongest open-weight model families by mid-2026 came from China, while describing the Western field as thinner after Meta’s open-model effort slowed. That comparison remains dependent on the benchmarks and definition of openness being used.
Open-weight models provide downloadable parameters, but they are not always fully open-source systems with public training data and reproducible training methods. The report contrasted the Chinese leaders with Ai2’s Olmo 3, which it described as more fully open but behind them in raw benchmark capability. It also linked China’s efficiency push partly to US chip-export restrictions, an interpretation rather than a confirmed explanation for every release.
“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 Licensing Leave Gaps
It is not yet clear whether the eight-week pace can be sustained or whether future releases will retain similarly permissive licenses. Benchmark results may also change with testing methods, model updates and task selection; the supplied material cites BenchLM and Artificial Analysis but does not include complete underlying evaluations.
Some product claims also require independent confirmation, including Kimi K2.7-Code’s reported 30% reduction in reasoning-token use compared with K2.6. The legal and security treatment of downloaded weights varies by jurisdiction, while hosted Chinese APIs may be subject to Chinese data rules. Restrictions on the DeepSeek application do not automatically establish equivalent restrictions on locally deployed weights.
Laplink PCmover – Easy Migration of your Applications, Files and Settings from an Old PC to a New PC – Data Transfer Software with Optional Super Speed USB 3.0 Cable – Business Standard, 10 Licenses
Versatile Licensing Options: PCmover Business offers flexible licensing with Standard License tiers for 1, 5, 10, or 25…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Deployers Face Fresh Model Choices
Developers and enterprise buyers will now test the four models on real workloads, hardware costs and data controls, rather than relying only on composite scores. The next markers will be independent benchmark replication, adoption figures, updated licensing terms and any government rules separating hosted services, consumer applications and locally operated model weights.
Key Questions
Which four models were released?
The reported sequence includes DeepSeek V4, MiniMax M3, Kimi K2.7-Code and GLM-5.2, released from April 24 through mid-June 2026.
Are these models open-source?
They are described as open-weight models, meaning their trained parameters can be downloaded. That does not necessarily mean their training data, code and methods are fully public or reproducible.
How close are they to proprietary models?
BenchLM’s July composite placed DeepSeek V4 Pro six points behind its proprietary leader. That result reflects one benchmark methodology, not a universal capability measure.
Can European organizations deploy them locally?
Downloadable weights can support local or on-premises deployment, subject to licensing, hardware and organizational rules. Using a hosted Chinese API creates different data-governance issues from running weights on controlled infrastructure.
Will the current release pace continue?
No continuation has been confirmed. Future cadence will depend on laboratory plans, computing access, commercial strategy and whether current licensing policies remain in place.
Source: Thorsten Meyer AI
