OPEN WEIGHTS · DATA RESIDENCY
On July 26, Moonshot AI released Kimi K3's open weights a day ahead of schedule: 2.8 trillion total parameters, 104 billion active, the first open model to cross the 3-trillion mark, ranking third on Artificial Analysis's Intelligence Index behind Fable 5 and GPT-5.6 Sol Max and first on Frontend Code Arena. The benchmark is not the load-bearing question for a small studio. Where the request physically executes is. Same model weights, four ways to call them, and one of them still lands in mainland China even when the front door reads American.
Same Kimi K3 weights, four call paths
Pick what kind of task you are routing, then see which providers clear it.
Executes on Moonshot's own infrastructure, mainland China.
Same PRC-hosting profile flagged on this radar July 9 for Kimi K2.7 Code, under China's National Intelligence Law.
Router is US-based, but currently forwards every request to Moonshot's own hosted INT4 endpoint.
A US front door does not change where inference runs. Same PRC hop as calling Moonshot directly.
Hosts the open weights on its own US-based serverless infrastructure with Zero Data Retention.
The request never leaves Fireworks. No PRC hop, and nothing retained after the call.
Runs entirely on hardware you control.
Zero third-party exposure, but the 594GB MXFP4 weights need roughly 8x H100 80GB minimum. Not something a solo studio provisions.
For client-confidential or NDA work, only Fireworks and self-hosting clear it. A US company name on the router is not the same as US-based inference.
What Kimi K3 actually benchmarks at
#3
Intelligence Index
behind Fable 5, GPT-5.6 Sol Max
#1
Frontend Code Arena
highest of any model
88.3
Terminal-Bench 2.1
agentic terminal tasks
1M
Context window
tokens, native vision
2.8 trillion total parameters, 104 billion active per token, MoE with 896 experts and 16 selected per token. Figures read directly from Moonshot's own README and technical report on GitHub.
WHY THIS MATTERS FOR CLIENT WORK
This radar flagged the same caveat back on July 9, when GitHub first put Moonshot's Kimi K2.7 Code in the Copilot model picker: Moonshot is a PRC company under China's National Intelligence Law, so anything routed to their own servers carries that exposure regardless of how the request got there. Kimi K3 being open-weight is what actually resolves it. Fireworks hosts the same weights on its own US infrastructure with Zero Data Retention, so near-frontier intelligence at open-weight pricing stops being a tradeoff against data residency once you pick the right provider, not the first one that shows up in a model picker.
Built 29 July 2026 · pattern sourced from Moonshot AI's Kimi K3 release