TL;DR
Four Chinese laboratories released 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 schedule, permissive licensing and lower hosted prices could accelerate local AI deployment, although benchmark limits, data rules and future licensing policies create uncertainty.
Four Chinese AI laboratories released frontier-class open-weight models in roughly eight weeks, according to a July 13 report from Thorsten Meyer AI, increasing competitive pressure on Western developers through lower hosted prices, downloadable weights and a faster upgrade cycle.
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. Most were issued under MIT or modified-MIT terms.
DeepSeek V4 uses a mixture-of-experts architecture with 1.6 trillion total parameters and 49 billion active parameters per pass, plus a one-million-token context window. MiniMax M3 combines a similarly long context with native multimodal support. Moonshot says Kimi K2.7-Code uses about 30% fewer reasoning tokens than K2.6 during agent runs.
Thorsten Meyer AI reported that hosted access to the Chinese models costs roughly five to 30 times less than Western frontier APIs. Its cited July BenchLM snapshot scored DeepSeek V4 Pro at 87, six points behind a proprietary leader at 93. The report cautioned that the figures represent one benchmark composite, rather than a final measure of model quality.
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.
A Faster Open-Model Cycle
The releases suggest that high-capability open-weight models are now being refreshed on a weeks-long schedule. That pace can lower the cost of experimentation and give businesses more frequent alternatives to closed commercial APIs.
For European organizations pursuing local or sovereign AI, downloadable models and permissive licenses make on-premises deployment more economically practical. The depth of the Chinese field also matters: DeepSeek, Z.ai, Moonshot and Alibaba now offer competing model families rather than leaving the market dependent on a single laboratory.
Waveshare ESP32-S3 1.43inch AMOLED Round Touch Display Development Board, 466 × 466, Supports AI Speech Interaction and Online LLM Platforms, Dual-core Processsor and 240MHz Frenquency, NO BAT
- Processor: Dual-core ESP32-S3-PICO-1-N8R8, 240MHz
- Connectivity: Wi-Fi 2.4GHz and Bluetooth 5 (LE)
- Memory: 8MB PSRAM and Flash, 512KB SRAM, 384KB ROM
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
China Builds Open-Weight Depth
Earlier competition around Chinese open models centered heavily on DeepSeek. The July report describes a broader field in which four Chinese developers occupy the upper tier, each emphasizing a different advantage: DeepSeek on price, Z.ai on benchmark performance, Moonshot on agent workloads and Alibaba’s Qwen on model variety and self-hosting.
The report says four of the five leading open-weight families now come from Chinese laboratories. It contrasts that expansion with a thinner Western field, while distinguishing open-weight releases from fully open-source systems that also publish training data or broader development materials.
“The cadence is the signal.”
— Thorsten Meyer AI, July 13 market report
Multimodal AI Systems Engineering: Building Production Vision-Language Models, Document AI, and Cross-Modal Retrieval Pipelines (Production AI Engineering Series)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Benchmarks and Access Carry Limits
It is not yet clear whether the release pace can be maintained or whether later models will retain permissive license terms. Beijing’s export posture, corporate strategy and access to advanced computing hardware could alter the schedule.
Benchmark rankings also do not establish performance across every workload. The supplied report cites BenchLM and Artificial Analysis, but results may vary in coding, reasoning, multilingual use and long-running agents. Independent testing will be needed to verify vendor efficiency claims, including Moonshot’s reported 30% token reduction.
Deployment rules remain another dividing line. Downloaded weights can run locally, while prompts sent to hosted Chinese APIs may be subject to Chinese data law. The report also says some Western agencies and enterprises reject Chinese-origin models, even where local use is technically possible.
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.
Independent Testing Moves to Deployment
Model evaluators and enterprise teams will now test the four releases across real production workloads, security controls and total operating costs. Their results will show whether benchmark gains translate into dependable use outside controlled tests.
Developers will also watch the next release cycle for evidence that the current pace and licensing terms can persist. European buyers face a parallel decision between local deployment, Chinese hosted services and Western proprietary APIs.
Building MCP Servers for AI Agents: Scalable Architecture Patterns, Security Design, and Production-Ready AI Infrastructure for Large Language Models
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 identifies DeepSeek V4, MiniMax M3, Moonshot AI’s Kimi K2.7-Code and Z.ai’s GLM-5.2.
Are these models open source?
They are described as open-weight models, meaning their trained weights can be downloaded. That does not always include the training data, code and documentation associated with fully open-source development.
How close are they to proprietary systems?
In the cited BenchLM snapshot, DeepSeek V4 Pro scored 87, compared with 93 for the proprietary leader. That six-point gap reflects one composite ranking and may differ by task.
Why are European organizations paying attention?
Long context windows, downloadable weights and lower prices can support local AI infrastructure. Organizations still need to review licensing, security, model origin and data-governance requirements.
Source: Thorsten Meyer AI
