TL;DR
Chinese AI labs released four frontier-class open-weight models between April 24 and mid-June 2026, marking a faster release cycle for downloadable systems. Benchmark and pricing data suggest Chinese models are narrowing the capability gap with closed rivals, though licensing, data governance and geopolitical risks remain.
Chinese AI laboratories released four frontier-class open-weight models between April 24 and mid-June 2026, compressing a major product cycle into roughly eight weeks. DeepSeek, MiniMax, Moonshot AI and Z.ai made their new systems downloadable, giving developers more high-capability alternatives to closed Western models while creating new questions about licensing, governance and dependence on Chinese technology.
DeepSeek V4, offered in Pro and Flash versions, arrived on April 24. According to Thorsten Meyer AI’s market review, the mixture-of-experts model contains 1.6 trillion total parameters but activates 49 billion for each pass, supports a one-million-token context window and uses the MIT license. Its hosted service also pushed down the price floor for high-end model access, the review said.
MiniMax M3 followed on June 1 with native multimodal capabilities, a one-million-token context window and a modified MIT license. Moonshot AI released Kimi K2.7-Code around June 13, positioning it as a specialist for agent-based coding tasks. The company says it uses about 30% fewer reasoning tokens than Kimi K2.6, a performance claim tied to that predecessor comparison.
Z.ai released GLM-5.2 between June 13 and June 16 under an MIT license. The 753-billion-parameter mixture-of-experts model ranked as the leading open-weight system on the Artificial Analysis index cited by Thorsten Meyer AI. The four releases were downloadable, while hosted access was reported to cost five to 30 times less than Western frontier APIs, depending on the model and service compared.
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.
As an affiliate, we earn on qualifying purchases.
Open Models Reach a Faster Cycle
The release sequence suggests that open-weight capability is advancing every few weeks, rather than through isolated annual launches. For businesses that can operate models locally, frequent improvements and permissive licenses can reduce the cost of on-premises AI deployments and lessen reliance on a single commercial API provider.
BenchLM’s July 2026 composite gave DeepSeek V4 Pro a score of 87, six points below the proprietary leader at 93. GLM-5.1 scored 83, Kimi K2.6 scored 81 and Qwen 3.5 397B scored 79. The figures support the view that several Chinese model families occupy the upper tier, but BenchLM is one benchmark provider and its composite is not a definitive measure of performance across every workload.
The breadth of the field also matters. Thorsten Meyer AI reported that four of the five strongest open-weight families now come from Chinese laboratories: DeepSeek, Z.ai, Moonshot AI and Alibaba. Their products target different needs, including low-cost inference, agent stability, long context and self-hosting, giving developers more choices than a ranking alone indicates.
Four Labs Build Distinct Positions
China’s open-model market is no longer centered on one leading developer. DeepSeek competes heavily on price, Z.ai targets benchmark performance, Moonshot AI focuses on long-running agents and Alibaba’s Qwen family spans many model sizes, including smaller variants suitable for local hardware.
The supplied market review contrasts that depth with a thinner Western field. It says Meta’s flagship open-model effort has lost momentum, while Ai2’s Olmo 3 remains fully open-source but trails the Chinese leaders on broad capability measures. Those comparisons depend on benchmark selection, model definitions and tested versions, all of which can change the ordering.
“The cadence is the signal.”
— Thorsten Meyer AI
Benchmark and Policy Risks Persist
It is not yet clear whether the current release pace will continue or whether future models will retain similarly permissive licenses. MIT-class terms make current weights easier to adopt, but later releases could use different conditions, and government export policy could also change.
Performance comparisons remain provisional. BenchLM’s six-point gap is based on a July composite score, while Artificial Analysis uses a separate methodology. Neither ranking establishes that one model will perform best for every coding, reasoning, multilingual or agent task. The claimed five-to-30-fold API price advantage also depends on token usage, service tier and the Western product selected for comparison.
Data governance presents another divide. Downloaded weights can run locally, but prompts sent to hosted Chinese APIs may fall under Chinese data law. Some Western agencies and regulated organizations restrict Chinese-origin applications or models. The supplied material says U.S. federal agencies have barred the DeepSeek app on government devices, while downloadable model weights remain legal and widely used; individual organizations may impose broader rules.
Deployers Will Test Cost and Control
Developers and enterprise buyers will now test the four releases against their own workloads, hardware and compliance rules. Independent evaluations of GLM-5.2, Kimi K2.7-Code, MiniMax M3 and DeepSeek V4 will help establish whether their benchmark positions translate into reliable production performance.
The next signal will be whether Chinese laboratories maintain a weeks-long model refresh cycle and continue offering downloadable weights under broad licenses. Buyers planning local deployments will also watch for license changes, export restrictions and data-handling guidance before making long-term infrastructure commitments.
Key Questions
Which four models were released?
The releases were DeepSeek V4 on April 24, MiniMax M3 on June 1, Kimi K2.7-Code around June 13 and GLM-5.2 between June 13 and June 16, 2026.
Are these models open-source?
They are described as open-weight models, meaning their trained weights can be downloaded. That does not always mean the training data, code and full development process are open. Most use MIT or modified-MIT terms, according to the supplied review.
How close are they to leading closed models?
BenchLM’s July composite placed DeepSeek V4 Pro at 87, compared with 93 for its proprietary leader. That six-point difference is one benchmark snapshot and may not reflect results on a specific task.
Why are European organizations interested?
Downloadable models can support local or sovereign AI deployments, keeping data within infrastructure controlled by the operator. Lower hosted prices and permissive licenses may also reduce costs, though Chinese-origin technology can trigger procurement restrictions.
What are the main adoption risks?
The main risks include changing license terms, export policy and uneven benchmark results. Organizations using hosted Chinese services must also examine data location, applicable law and internal compliance requirements.
Source: Thorsten Meyer AI
