Kimi K3, live on PrivateMind
Kimi K3 is now serving on PrivateMind behind the same gateway and audit trail as everything else we run. What it took to productionise, and what the first few hours looked like.
Archive
Archive
Hopper vs Blackwell: picking the right chip for finance
Blackwell is faster on paper. Whether it is faster for your workload depends on batch shape, memory bandwidth and what you actually run.
Hugging Face and the trust problem
What a supply-chain incident at a model host says about where your weights live and who can reach them.
Kimi K3 and the open-weight frontier
Moonshot AI released a 2.8-trillion-parameter open-weight model that trades blows with the closed frontier. The gap is closing in steps, and the steps are getting larger.
From London to New York: Three themes that dominated every room
Three themes came up in every room on both sides of the Atlantic: control of data, the cost of inference, and who you are actually trusting.
Inside the gateway: why one choke point
One process sees every prompt, every token and every key. Why we built it that way, and what it buys us.
GLM-5.2 and the open source frontier
GLM-5.2 lands within a point of the frontier on long-horizon coding. What it means for the open-source race, and the three things it unlocks: cost, performance, sovereignty.
Sovereign by default
Sovereignty is not a deployment option you bolt on later. It is a property of where the weights sit and who can reach them.
Private AI, in the open
We run AI nobody else gets to see into, and we write down how we build it where everyone can. Why this blog exists, and what it will and won't be.