China just dropped a model bigger than GPT-4 and handed out the weights, but the fine print says "open" isn't what it used to be.
The Summary
- Moonshot AI released the full 2.8 trillion-parameter weights for Kimi K3, plus a 47-page technical report and the infrastructure to self-host it
- The model activates 104 billion parameters from 896 experts, runs a million-token context window, and hits frontier benchmarks
- It's "open" for commercial use, but the custom license adds obligations on larger enterprises that Apache 2.0 and MIT don't have
- This is the first 3T-class model with full weights available for download and modification
The Signal
Moonshot AI just made the biggest model release of 2026 that most people will sleep on. Kimi K3's full weights are now available, and it's not just the parameters that matter. It's what comes with them: inference infrastructure, optimized attention kernels, MoE communication libraries, and deployment components. This isn't a research toy. It's a production-ready package for enterprises that want to run frontier AI on their own metal.
The architecture is worth understanding. Kimi K3 uses a Mixture-of-Experts design with 896 experts, activating just 104 billion parameters per forward pass. That's the trick: you get 2.8 trillion parameters worth of knowledge density, but you only pay compute costs for 104 billion at inference time. Add native multimodal reasoning and a million-token context window, and you've got a model that can process entire codebases, legal documents, or research papers in a single pass.
"This is the world's first open 3T-class model with full weights available for download and modification."
But here's where it gets interesting:
- The license grants commercial rights, modification rights, and offline deployment rights
- It also adds obligations on larger companies that traditional open-source licenses skip
- Those obligations aren't disclosed in the VentureBeat piece, but they exist in a custom 47-page usage agreement
- This is a new licensing pattern emerging from Chinese AI labs that splits the difference between "closed" and "truly open"
The technical innovations matter too. Kimi Delta Attention, Attention Residuals, and Stable LatentMoE are architectural choices that make this model trainable and deployable at scale. Moonshot is releasing implementation support for vLLM and SGLang, which means the model plugs into existing inference stacks. That's not a given with models this large.
The real question is what "obligations on larger companies" means in practice. Is it revenue sharing above a certain threshold? Disclosure requirements? Geographic restrictions? The VentureBeat story doesn't say, but it flags the issue clearly enough that enterprise buyers should read every page of that custom license before they spin up a cluster.
The Implication
If you're building agents that need deep reasoning, long context, or offline operation, Kimi K3 just became the most credible alternative to closed frontier models. But don't assume "open weights" means "no strings attached." The licensing innovation here is as important as the technical one. Chinese AI labs are writing a new playbook for model distribution, one where "open" includes commercial rights but stops short of Apache 2.0 freedom.
Watch for other labs to copy this pattern. The era of purely altruistic open-source AI might be ending. The era of strategically open AI is here.