TL;DR
Four Chinese laboratories released downloadable, high-capability open-weight AI models between April 24 and mid-June 2026. July benchmark snapshots place several Chinese model families near leading proprietary systems, although pricing, performance and licensing comparisons require independent validation.
Four Chinese AI laboratories released downloadable open-weight models within roughly eight weeks, from DeepSeek V4 on April 24 through mid-June releases from Moonshot AI and Z.ai. The rapid schedule matters because July benchmark snapshots place several Chinese model families near the proprietary frontier, while hosted access reportedly costs far less than leading Western APIs.
The sequence began with DeepSeek V4 Pro and Flash, described as mixture-of-experts models with 1.6 trillion total parameters, 49 billion active parameters and a one-million-token context window. MiniMax M3 followed on June 1 with native multimodal support and a modified MIT-style license, according to the Thorsten Meyer AI market report.
Moonshot AI released Kimi K2.7-Code around June 13, targeting long-running coding agents. The company’s reported efficiency comparison says it uses about 30% fewer reasoning tokens than K2.6. Z.ai released GLM-5.2 in mid-June; the report describes it as a 753-billion-parameter mixture-of-experts model distributed under the MIT license.
All four releases are described as downloadable open-weight models, though that term does not mean their training data, full development process or every software component is public. The report says most carry MIT or modified-MIT terms and estimates hosted prices at five to 30 times below Western frontier APIs. That cost range is a market comparison, not a fixed ratio across every workload.
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.
![Claude AI for Beginners Bible: [5 in 1] The Ultimate Guide to Automate Your Work, Save Hours Every Week, and Use AI for Real-World Results](https://m.media-amazon.com/images/I/415+fSJacsL._SL500_.jpg)
Claude AI for Beginners Bible: [5 in 1] The Ultimate Guide to Automate Your Work, Save Hours Every Week, and Use AI for Real-World Results
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
China Accelerates the Open-Model Cycle
The main development is the frequency of competitive releases, not any single benchmark result. A weeks-long update cycle gives developers more frequent options for replacing or supplementing proprietary APIs, while long context windows and sparse architectures may reduce the cost of local document processing, coding agents and other demanding applications.
The trend has particular relevance for European local-first deployments. Downloadable weights can be run within an organization’s own infrastructure, allowing prompts and outputs to remain under the operator’s technical control. That can support data-governance goals, but model origin, license terms and deployment controls may still affect procurement decisions in government and regulated industries.
The July snapshot also suggests that Chinese open-weight development has become several laboratories deep. BenchLM placed DeepSeek V4 Pro at 87, six points behind a proprietary leader at 93, while earlier models from Z.ai, Moonshot and Alibaba scored 83, 81 and 79. Those figures come from one composite benchmark tracker and should not be treated as universal measures of quality.
AI Systems Performance Engineering: Optimizing Model Training and Inference Workloads with GPUs, CUDA, and PyTorch
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Four Laboratories Now Compete Near the Top
Chinese open-weight development was previously associated most strongly with DeepSeek and Alibaba’s Qwen family. The spring releases add visible competition from MiniMax, Moonshot AI and Z.ai, with different priorities spanning low-cost inference, multimodal input, coding agents and broad benchmark performance.
The Thorsten Meyer AI report says four of the five strongest open-weight families were produced by Chinese laboratories as of July 2026. It contrasts that depth with a thinner Western field, while identifying Ai2’s Olmo line as a more fully open alternative. The comparison depends on how openness and capability are defined, since open weights, open-source code and disclosed training data are separate attributes.
“The individual releases got their headlines. The cadence didn’t — and the cadence is the signal.”
— Thorsten Meyer AI, AI Dispatch report
LAFVIN ESP32-S3 1.69" LCD Development Board with Camera, AI Vision Voice Development Kit, Programmable IoT Board with Mic Speaker for STEM Education
【Abundant Core Computing Power】 Powered by the ESP32-S3 microcontroller and equipped with a large-capacity memory configuration of 16MB…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Benchmarks and Future Licenses Remain Fluid
It is not yet clear how all four models compare under independent, workload-specific testing. The reported benchmark positions, token savings and API price differences use different methodologies and may change with model updates. The durability of permissive licensing is also unknown, as are any future Chinese export restrictions or policy changes affecting model distribution. Hosted Chinese services carry separate data-jurisdiction questions that do not automatically apply to locally operated 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.
Independent Testing Moves to Deployment
Developers and procurement teams are expected to test the releases against real coding, multimodal and long-context workloads, rather than relying only on composite scores. Attention will also turn to license stability, hardware requirements and data handling. Further model releases or benchmark updates could quickly alter the July rankings, making the current table a dated snapshot rather than a settled order.
Key Questions
Which four models were released?
The reported sequence includes 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 fully open source?
They are described as open-weight and downloadable. That allows local operation, but it does not necessarily include public training data, training code or full development records.
Are Chinese open models now equal to closed frontier systems?
BenchLM’s July composite placed DeepSeek V4 Pro six points behind its proprietary leader. That indicates a smaller reported gap, but one benchmark cannot establish parity across every task.
Why are European organizations watching these releases?
The models offer local deployment and lower reported costs, which may help organizations keep sensitive processing on their own infrastructure. Procurement rules, model origin and hosted-service data jurisdiction may still limit adoption.
Could the licenses change?
Existing releases remain governed by their published terms, but future versions could use different licenses. The long-term availability of permissively licensed Chinese models has not been guaranteed.
Source: Thorsten Meyer AI
