TL;DR
Four Chinese laboratories released high-end open-weight AI models between April 24 and mid-June 2026, marking a faster release cycle for downloadable systems. July benchmark data places the strongest Chinese model six points behind a proprietary leader, but vendor specifications, licensing policies and regulatory exposure require separate review.
Four Chinese AI laboratories released high-end open-weight models between April 24 and mid-June 2026, compressing what was once an annual upgrade cycle into roughly eight weeks. The arrival of DeepSeek V4, MiniMax M3, Kimi K2.7-Code and GLM-5.2 matters because downloadable models are approaching leading proprietary systems on broad benchmarks while offering lower hosted prices and more local deployment options.
DeepSeek V4 Pro and Flash arrived 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 Thorsten Meyer AI dispatch. The report describes every model as downloadable and says most carry MIT or modified-MIT licensing.
The published specifications point to different technical priorities. DeepSeek says V4 uses a 1.6-trillion-parameter mixture-of-experts design that activates 49 billion parameters per pass and supports a one-million-token context window. MiniMax promotes M3 as a low-cost, multimodal model with the same context length. Moonshot says Kimi K2.7-Code uses about 30% fewer reasoning tokens than K2.6 during agent tasks. These specifications are vendor claims and have not been independently validated in the supplied material.
In BenchLM’s July 2026 composite, DeepSeek V4 Pro scored 87, compared with 93 for the unnamed proprietary leader. GLM-5.1 scored 83, Kimi K2.6 scored 81 and Qwen 3.5 397B scored 79. The results indicate several competitive Chinese model families, although they come from one benchmark provider and do not establish performance across every workload. Artificial Analysis separately ranked GLM-5.2 as its leading open-weight model, according to the dispatch.
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.
![MixPad Free Multitrack Recording Studio and Music Mixing Software [Download]](https://m.media-amazon.com/images/I/71ltIxIuz1L._SL500_.jpg)
MixPad Free Multitrack Recording Studio and Music Mixing Software [Download]
Create a mix using audio, music and voice tracks and recordings.
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
China Accelerates the Open Model Cycle
The release sequence suggests that open-weight capability is being refreshed within weeks, led by several Chinese laboratories rather than a single company. That pace can lower the cost of testing new systems and gives developers more choices among DeepSeek, Z.ai, Moonshot AI and Alibaba, each of which emphasizes a different mix of price, scale, agent performance and self-hosting.
For European companies pursuing local or sovereign AI, downloadable weights and permissive licenses can make on-premises deployment more practical. The dispatch estimates that hosted Chinese APIs cost five to 30 times less than Western frontier services, though exact savings depend on usage patterns, infrastructure and contract terms. Local deployment can also keep prompts away from an external API, but it does not by itself settle questions about model provenance, security review or regulatory approval.
Yahboom Raspberry Pi 5 ROS2 Robot Car 360°Movement, AI Vision & Tracking, Integrated Multimodal Large AI Model OpenRouter, AI Voice Interaction (Superior Without RPi5)
【Powerful control system】RaspberryPi 5 has made breakthroughs in processor speed,multimedia performance,memory and connection.Based on the RaspberryPi 5 main…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Chinese open-weight market was previously associated most closely with DeepSeek and Alibaba’s Qwen family. The July rankings cited by the dispatch show four laboratories in the upper tier: DeepSeek, Z.ai, Moonshot AI and Alibaba. Qwen remains the broadest family in the group, with smaller variants suited to single-GPU systems, while DeepSeek has placed greater emphasis on low pricing and sparse model architecture.
The comparison also reflects a distinction often blurred in model coverage. Open-weight models provide downloadable parameters, but they are not automatically open-source systems with public training data, complete training code and reproducible development records. Licensing varies by release, and benchmark proximity does not establish operational equivalence with proprietary models in reliability, safety, tool use or specialized tasks.
“That’s not a wave. That’s a production line.”
— Thorsten Meyer AI dispatch
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.
Benchmarks and Access Policies May Shift
It is not yet clear whether the eight-week release pace can be maintained or whether later models will keep similarly permissive licenses. Model creators may change usage conditions, distribution policies or export availability, leaving companies that depend on one family exposed to commercial and geopolitical policy changes.
The broad performance claims also need more independent testing. BenchLM’s scores are a single composite snapshot from July 2026, while Artificial Analysis uses a different methodology. Neither ranking alone answers how the models perform on a company’s own data, under sustained agent workloads or after security controls are applied. The supplied material also does not provide independent confirmation of pricing ratios or Moonshot’s claimed token reduction.
Data governance depends on the deployment method. Downloaded weights can run locally, while hosted Chinese APIs process prompts under Chinese law. The dispatch says Western agencies and regulated organizations may reject Chinese-origin models and reports restrictions on the DeepSeek app on US government devices. The exact reach of those restrictions, and their treatment of locally deployed model weights, requires jurisdiction-specific verification.
EarlySincere 4K AI Smart Glasses with Blue-Light Blocking Photochromic Lenses, Integrated ChatGPT AI Model, Real-Time Translation, HD Video Recording, Long Battery Life, Bluetooth 5.4 Wireless
👓【𝟒𝐊 𝐕𝐈𝐃𝐄𝐎 𝐑𝐄𝐂𝐎𝐑𝐃𝐈𝐍𝐆 𝐆𝐋𝐀𝐒𝐒𝐄𝐒 𝐖𝐈𝐓𝐇 𝐄𝐈𝐒】 This advanced wearable camera empowers you to capture the wonders of the…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Independent Testing Becomes the Next Test
Developers and procurement teams will now test the four releases against real workloads, security requirements and total operating costs. Upcoming independent evaluations should show whether their benchmark positions carry over to coding agents, multimodal tasks, long-context retrieval and sustained production use.
Organizations adopting the models are also likely to track license changes, export policy and API data handling. A practical response is to preserve portability across several model families rather than depend on one provider. The next major signal will be whether Chinese laboratories maintain the weeks-long release cadence and whether Western open-weight developers answer with comparable systems.
Key Questions
Which four Chinese models were released?
The sequence comprised DeepSeek V4 on April 24, MiniMax M3 on June 1, Moonshot AI’s Kimi K2.7-Code around June 13 and Z.ai’s GLM-5.2 in mid-June 2026.
Are these models open source?
They are described as open-weight models, meaning their parameters can be downloaded. That does not necessarily include training data, full training code or reproducible documentation, so they should not all be treated as fully open-source systems.
How close are they to proprietary AI models?
BenchLM’s July composite placed DeepSeek V4 Pro at 87, six points behind a proprietary leader at 93. This shows proximity on one benchmark composite, not equal performance across all tasks.
Can European organizations deploy them locally?
Downloadable weights can support local deployment, subject to hardware, license, security and regulatory reviews. Using a hosted Chinese API presents separate data-governance questions because prompts leave the organization’s infrastructure.
Why are the releases drawing attention now?
The main development is the concentration of four high-end releases in eight weeks. Combined with low advertised API prices and permissive licensing, that cadence suggests Chinese laboratories now drive much of the open-weight market’s pace.
Source: Thorsten Meyer AI
