In the back half of July, Chinese and independent labs shipped trillion-parameter models you can run on your own hardware under MIT and Apache licenses. For firms that stayed off frontier AI over data residency, the blocker just moved.
For two years, "open weights" meant a smaller model you settled for. In the back half of July 2026, that stopped being true. A cluster of labs shipped trillion-parameter models you can download, and the licenses are the permissive kind.
Start with DeepSeek V4. Per the model cards on Hugging Face, the V4 series ships two Mixture-of-Experts models, V4-Pro at 1.6 trillion parameters (49 billion active) and V4-Flash at 284 billion (13 billion active), both with a one-million-token context window, both under the MIT license. MarkTechPost reports the weights were public from day one, and the models now sit on Dell's Hugging Face Enterprise Hub, which is about as enterprise a signal as an open model gets.
Then Kimi K3. Moonshot AI's new model is a 2.8-trillion-parameter Mixture-of-Experts system with native image understanding, and Nathan Lambert's Interconnects reports the open weights land July 27. Moonshot is not a side project. Per reporting from Qz and TechNode, its annual recurring revenue reached $300 million in June after K3 shipped, and it is preparing a pre-IPO round targeting a valuation as high as $50 billion ahead of a Hong Kong listing.
Why a regulated firm should care
The MIT and Apache licenses on these models permit commercial use, modification, and resale, with no revenue thresholds and no attribution clause. Open weights mean the model runs inside your environment instead of someone else's API. Together those two facts answer the question that has kept a lot of law firms, banks, and health systems off frontier AI: where does the data go. Now it can stay home.
Here is the trade you are making.
| Managed API (frontier lab) | Self-hosted open weights (MIT / Apache) |
| The vendor holds the weights and the uptime | You hold the weights and the uptime |
| Your data leaves your environment | Your data stays inside it |
| Per-token pricing, no hardware to run | Hardware and operations you own |
| Licensing terms set by the vendor | Commercial use with no revenue thresholds |
| Patching and safety handled for you | Patching, evaluations, and guardrails are yours |
The honest part
Open weights hand you data sovereignty and a bill. Once the model lives on your hardware, you own the hosting cost, the security patching, the evaluations, and the guardrails a managed vendor used to run for you. You also own two questions a download does not answer. First, whether a Chinese-trained model belongs in your stack, given your clients and your regulators. Second, whether the model is safe to point at privileged client data, which an MIT license says nothing about. You pay nothing for the license. You pay for everything after it.
Your move: if data residency was the reason your firm stayed off frontier AI, that reason is gone. Pick one contained use case, stand up one open-weight model inside your own environment, and run a real evaluation before anyone says the word production. The download is the easy part.
AI, Properly is the Ayvede field report. Advise. Vet. Deliver. Ask a question in writing.