TL;DR
Chinese laboratories released four frontier-class open-weight AI models between April 24 and mid-June 2026. The rapid cycle, low hosted prices and permissive licenses strengthen the case for self-hosted AI, but benchmark limits, regulatory exposure and future licensing policies remain concerns.
Chinese AI laboratories released four frontier-class open-weight models between April 24 and mid-June 2026, a roughly eight-week production cycle that included systems from DeepSeek, MiniMax, Moonshot AI and Z.ai. The releases matter because their downloadable weights, permissive licenses and low hosted prices are narrowing the practical gap between self-hosted AI and proprietary services.
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 within days, according to the release timeline compiled by Thorsten Meyer AI.
DeepSeek V4 uses a mixture-of-experts architecture with 1.6 trillion total parameters and 49 billion activated for each pass, alongside a one-million-token context window. MiniMax M3 combines a similarly long context window with native multimodal support. Kimi K2.7-Code targets agent-based coding work and reportedly uses about 30% fewer reasoning tokens than K2.6 during long runs.
Z.ai’s GLM-5.2 has 753 billion parameters in a mixture-of-experts design and carries an MIT license, according to the source material. The report says most models in the release group use MIT or modified-MIT terms, while hosted access costs an estimated five to 30 times less than Western frontier APIs. Exact price comparisons depend on workload, token mix and provider discounts.
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.
Open AI Moves to Weekly Cycles
The cadence changes planning for companies seeking local or sovereign AI infrastructure. Model capability that once appeared to advance annually is now being refreshed within weeks, while permissive licenses and long context windows make advanced systems available for adaptation, private deployment and lower-cost experimentation.
The July 2026 BenchLM composite places DeepSeek V4 Pro at 87 points, six behind a proprietary leader scoring 93. GLM-5.1 scored 83, Kimi K2.6 scored 81 and Qwen 3.5 397B scored 79. Those results suggest several Chinese model families occupy the upper open-weight tier, rather than one laboratory carrying the field. BenchLM remains a single benchmark composite and does not settle performance across every task.
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.
Four Labs Build Competitive Depth
The release run reflects distinct technical and commercial strategies. DeepSeek competes heavily on price; Z.ai targets broad model capability; Moonshot AI focuses on long-running agents; and Alibaba’s Qwen family offers a wide range of model sizes, including versions suited to single-GPU use.
Thorsten Meyer AI reports that four of the five strongest open-weight families now come from Chinese laboratories. Western developers still publish open and open-weight models, including Ai2’s Olmo family, but the source argues that the highest benchmark scores and fastest release rhythm have shifted toward China. That conclusion is based on current rankings and could change with later releases or revised evaluations.
“The cadence is the signal.”
— Thorsten Meyer AI
LLM Systems Engineering: Training and Building Large Language Models – Engineering AI Models Through Fine-Tuning, Continued Pretraining, and From-Scratch Development
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Benchmarks and Dependency Risks Persist
It is not yet clear whether the current release pace can be sustained or whether later models will retain similarly permissive terms. An MIT license attached to one release does not bind a laboratory’s future products, and government export policy could also change.
Open weights and hosted APIs carry different risks. Downloaded models can run on private infrastructure, while prompts sent to Chinese-hosted services may fall under Chinese data rules. Some Western agencies and regulated organizations restrict Chinese-origin AI systems, but policies vary. The source material also links rapid efficiency gains partly to hardware constraints created by United States export controls; the relative influence of technical necessity, commercial competition and state strategy remains uncertain.
ESP32 Basic Starter Ai Chatbot Kit Development Board USB-C Dual Core Microcontroller Support AP/STA/AP+STA Compatible with Arduino IDE IoT for Beginners Engineers
- High-Performance Dual-Core CPU: Equipped with dual-core processor and Type-C USB
- Rich Peripheral Support: Includes SPI, LCD, Camera, UART, I2C, and more
- Wireless Connectivity: Supports Wi-Fi 2.4 GHz and Bluetooth 5 (LE)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
New Releases Will Test Durability
Developers will watch the next release cycle for independent benchmark replication, detailed safety evaluations, stable licensing and real-world agent performance. Buyers comparing local deployment with hosted access will also need to track data residency, hardware costs and provider terms. The next major Chinese or Western open-weight release will show whether this eight-week run marks a durable production rhythm or an unusually concentrated period.
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
- Flexible Licensing Tiers: 1, 5, 10, or 25 transfers
- Unlimited Transfer Attempts: One license per source-destination pair
- Suitable for Multiple PCs: Supports individual or multi-PC migrations
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
Which four models were released during the eight-week period?
The reported sequence comprised DeepSeek V4 on April 24, MiniMax M3 on June 1, Kimi K2.7-Code around June 13 and GLM-5.2 in mid-June 2026.
Are these models fully open source?
They are described as open-weight models, meaning their trained parameters can be downloaded. That is not always the same as full open source, which may also require training code, data details and reproducible methods.
Are Chinese open models now equal to proprietary leaders?
Not across every measure. BenchLM placed DeepSeek V4 Pro six points behind its proprietary leader in July, but benchmark results vary by task. Independent testing and deployment results remain necessary.
Why are these releases relevant to European organizations?
They offer lower-cost options for local deployment, which can support data-control goals. Organizations must still review license durability, model provenance, security and applicable regulation before adoption.
Source: Thorsten Meyer AI
