TL;DR
Chinese laboratories released four frontier-class open-weight AI models between April 24 and mid-June 2026. The rapid release cycle is lowering self-hosting costs, but benchmark limits, licensing durability and regulatory exposure remain unresolved.
Chinese AI laboratories released four frontier-class open-weight models in roughly eight weeks, from DeepSeek V4 on April 24 through the mid-June arrivals of Kimi K2.7-Code and GLM-5.2. The compressed schedule matters because it suggests that high-capability open models are now improving on a cycle measured in weeks, changing the cost and planning assumptions behind self-hosted AI.
The sequence began with DeepSeek V4 Pro and Flash, followed by MiniMax M3 on June 1, Moonshot AI’s Kimi K2.7-Code on June 13 and Z.ai’s GLM-5.2 within days of the Kimi release. Thorsten Meyer AI classified all four as frontier-class and downloadable, with most carrying MIT or similarly permissive licenses.
DeepSeek V4 uses a mixture-of-experts architecture with a reported 1.6 trillion total parameters, 49 billion active parameters per pass and a one-million-token context window. MiniMax M3 also offers a one-million-token context and native multimodal support. Moonshot positioned K2.7-Code for extended agent tasks, while Z.ai released GLM-5.2 as a 753-billion-parameter mixture-of-experts model under an MIT license.
The report says hosted versions of the Chinese models cost five to 30 times less than Western frontier APIs, depending on the comparison. Exact savings will vary with providers, token mix, workload and infrastructure. Downloadable weights can remove per-call API charges, but self-hosting still carries hardware, energy and operating costs.
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.
Ship Local AI with Ollama: Build 5 Real Projects on Your Own Hardware in 30 Days (Local AI Playbooks)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Open Models Gain Release Speed
The cadence gives developers and enterprises more frequent capability upgrades without binding every workload to a proprietary provider. For European organizations pursuing local or sovereign AI deployments, permissively licensed weights and long context windows can make on-premises systems more practical.
The July BenchLM composite cited by Thorsten Meyer AI placed DeepSeek V4 Pro at 87, six points behind a closed proprietary leader scoring 93. GLM-5.1 scored 83, Kimi K2.6 scored 81 and a Qwen 3.5 model scored 79. Those results point to depth across several Chinese laboratories, rather than reliance on a single supplier, although one composite ranking cannot settle overall model quality.
Four Labs Crowd the Upper Tier
The Chinese open-weight field now includes distinct model strategies. DeepSeek competes heavily on price; Z.ai targets benchmark leadership; Moonshot emphasizes long-running agent performance; and Alibaba’s Qwen family spans models small enough for more accessible self-hosting. Thorsten Meyer AI estimates that four of the five strongest open-weight families now come from Chinese laboratories.
The report contrasts that breadth with a thinner Western open-model field, citing Meta’s slower flagship release activity and lower raw benchmark results for Ai2’s fully open Olmo line. That comparison depends on definitions: open-weight models publish parameters, while fully open projects may also release training code, data details and other materials needed for reproduction.
“The cadence is the signal.”
— Thorsten Meyer AI, July 13 market report
Benchmarks and Access Carry Limits
Several parts of the market picture remain unsettled. BenchLM’s scores are a single composite snapshot, and the supplied material does not provide enough methodological detail to verify the claimed six-point frontier gap independently. Performance may also differ across coding, reasoning, multilingual, multimodal and agent workloads.
The durability of permissive licensing is also unknown. Future releases could use different terms, and government policy in China or importing countries could affect distribution. Hosted Chinese APIs may process data under Chinese legal jurisdiction, while locally deployed weights present a different technical and regulatory profile. Some Western agencies and enterprises may still reject Chinese-origin models regardless of where they run.
Deployers Face New Model Choices
Model buyers will next watch whether Chinese laboratories maintain the weeks-long release pace, whether independent evaluations confirm the reported gains and whether the newest models hold up in production. Organizations evaluating them will also need to compare hosted and self-managed deployments, review license terms release by release and test models against their own workloads.
The next milestone is broader independent testing of DeepSeek V4, Kimi K2.7-Code and GLM-5.2. Their real impact will depend less on launch-day rankings than on reliability, operating cost, security review and adoption across regulated environments.
Key Questions
Which four models were released?
The reported sequence comprises 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 are described as open-weight and downloadable, but that does not automatically mean fully open source. Access to training data, code and complete development records may vary by model.
How close are they to proprietary frontier models?
BenchLM’s July composite placed DeepSeek V4 Pro six points behind its leading closed model. That gap applies to one benchmark composite and does not prove equal performance across every task.
Why is the release pace important?
A shorter cycle means self-hosted AI can improve more frequently, potentially lowering the capability and cost gap with proprietary services. It also forces buyers to make faster infrastructure and model-selection decisions.
What risks do European users face?
Hosted services can create data-jurisdiction and compliance concerns. Local deployment limits exposure to an external API, but organizations must still review licensing, security, origin policies and future availability.
Source: Thorsten Meyer AI
