TL;DR
Chinese labs released four frontier-class open-weight AI models between April 24 and mid-June 2026, according to a July 13 market report from Thorsten Meyer AI. The rapid cadence could lower self-hosting costs, but benchmark comparisons, data governance and the durability of permissive licensing remain unsettled.
Chinese AI laboratories released four frontier-class open-weight models between April 24 and mid-June 2026, according to a July 13 market report from Thorsten Meyer AI. The releases from DeepSeek, MiniMax, Moonshot AI and Z.ai compressed a high-end model cycle into roughly eight weeks, increasing pressure on Western providers while expanding options for organizations seeking lower-cost, self-hosted AI.
The sequence began with DeepSeek V4 Pro and Flash on April 24, followed by MiniMax M3 on June 1. Moonshot AI’s Kimi K2.7-Code arrived around June 13, while Z.ai released GLM-5.2 during the June 13-16 period, according to the report. All four were described as downloadable open-weight systems, with most carrying MIT or modified-MIT licenses.
DeepSeek V4 Pro was reported as a 1.6-trillion-parameter mixture-of-experts model that activates 49 billion parameters per pass and supports a one-million-token context. MiniMax M3 was presented as a low-cost, multimodal model with the same context length. Moonshot positioned Kimi K2.7-Code for long-running agent tasks, while GLM-5.2 was described as a 753-billion-parameter mixture-of-experts system.
BenchLM’s July composite gave DeepSeek V4 Pro a score of 87, six points behind a proprietary leader at 93. The same snapshot placed GLM-5.1 at 83, Kimi K2.6 at 81 and a Qwen 3.5 model at 79. These scores come from one benchmark tracker and do not establish performance 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.
Open Models Accelerate Deployment Choices
The release schedule suggests that high-capability open models are being refreshed within weeks, rather than through annual product cycles. For developers and enterprises, faster updates and permissive licenses may reduce the cost of running models on their own infrastructure. Thorsten Meyer AI estimated that hosted access to the Chinese systems was five to 30 times cheaper than Western frontier APIs, although the report did not provide a uniform workload comparison for that range.
The development is especially relevant for European organizations pursuing local or sovereign AI deployments. Downloadable weights can keep processing inside controlled infrastructure, but model origin, security review and procurement rules may still prevent adoption. Hosted Chinese services carry a separate concern because prompts may be processed under Chinese data law.
Server Room Temperature and Humidity Monitor for Data Centers,Pharmaceuticals Alongwith Factory Calibration Certificate Model: AI-RHTx-IOT (RHTx-IoT Hosting to Customer End (Without Hosting))
- Model and Connectivity: RHTx-IoT with SMS, Email, Cloud
- Measured Parameters: Temperature and Humidity
- Temperature Range and Accuracy: 0-50°C, ±0.5°C
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
China’s Open-Weight Bench Deepens
The report argues that China’s open-weight market is no longer dependent on a single leading laboratory. DeepSeek competes on price, Z.ai on benchmark performance, Moonshot on agent workloads and Alibaba’s Qwen family on a broad range of models, including versions suited to smaller hardware. BenchLM data cited in the report placed four Chinese model families near the top of its open-weight rankings.
Open-weight does not always mean fully open-source. Users may receive model weights and broad usage rights without obtaining the full training data, code or development record. The report identified Ai2’s Olmo 3 as a stronger example of full openness, while saying it trailed the Chinese leaders on raw benchmark capability.
“That’s not a wave. That’s a production line.”
— Thorsten Meyer AI, July 13 market report
Local AI Engineering with Ollama: Run, understand, customize, fine-tune, and build agentic apps on your own hardware
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Benchmarks and Licensing Remain Fluid
Several claims still require broader testing. It is not yet clear whether the new models deliver their reported gains across production coding, multimodal and agent workloads, or whether the claimed cost advantage persists once infrastructure, latency and support are included. The cited rankings are historical benchmark snapshots, not guarantees of future or real-world performance.
The durability of current licensing is also unknown. Future releases could use different terms, while government export policies or procurement restrictions could narrow access. The source says U.S. federal agencies have restricted the DeepSeek application on government devices, but distinguishes that action from the legal availability of downloadable weights. Organization-specific security and compliance decisions may go further.
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.
Independent Testing Moves to Center Stage
Attention will now shift to independent evaluations of the four releases, including tests of cost, reliability, security and hardware needs. Buyers will also watch whether Chinese laboratories maintain the same weeks-long release pace and permissive licensing. Enterprises weighing deployment will need separate reviews for downloaded models and hosted APIs because their data exposure and regulatory profiles differ.
Large Language Models – LLM and API's
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
Which four models were released?
The report identified DeepSeek V4, MiniMax M3, Moonshot AI’s Kimi K2.7-Code and Z.ai’s GLM-5.2.
Are all four models open-source?
They were described as open-weight and downloadable. That does not necessarily include full training data, source code or development records, so open-source status varies.
How close are they to proprietary frontier models?
On BenchLM’s July composite, DeepSeek V4 Pro scored 87 against 93 for the proprietary leader. That six-point difference applies only to that benchmark snapshot.
Why might European organizations use these models?
Downloadable weights can support local processing and potentially lower costs. Chinese origin, licensing, security policy and rules governing hosted services may still limit use in regulated or public-sector workloads.
Is the reported pricing advantage confirmed?
Thorsten Meyer AI reported hosted prices five to 30 times below Western frontier APIs. The available source does not show a standardized comparison covering usage patterns, infrastructure and service levels, so the range should be treated as a sourced market estimate.
Source: Thorsten Meyer AI
