hawkes.aiKey ideas
- The one idea — Trillion-scale MoE models are now downloadable — mostly from Chinese labs, mostly permissively licensed — a real shift in who ships frontier-scale weights.
- How it is shown — The open summits: Kimi's trillion-scale K-line and GLM as walkable peaks rivaling the sealed ones; the fits-by-total/streams-by-active reminder; the geography of who builds them.
- The trap to avoid — Assuming frontier-scale means American and closed — the largest downloadable models invert both assumptions.
- What it sets up — But the absolute ceiling is still sealed.
You can legally download trillion-parameter AI models today — free, permissive licenses. Mostly, they're not coming from Silicon Valley.
The one idea
Trillion-scale MoE models are now downloadable — mostly from Chinese labs, mostly permissively licensed — a real shift in who ships frontier-scale weights.
Last episode's open ridge grew summits nobody predicted: trillion-parameter models you can download, own, and run — rivaling the sealed peaks in scale. And their provenance rewrites the map. The biggest brains anyone can legally own come, overwhelmingly, from labs the early family tree barely mentioned. The open trillion era. The summits: the GLM line, Kimi's K-series, the DeepSeek giants — all mixture-of-experts, hundreds of billions to a full trillion total. And crucially, most ship under permissive seals — genuinely ownable, not just visible. A model with a trillion parameters and an open license would have sounded like a contradiction three years ago. It's now a category with several members. But recall episode one-twenty-three, because marketing won't. These fit by total, stream by active: a trillion-parameter open giant still demands rack-scale memory to host. Ownable is not laptop-runnable.
How it works — the demo
The open summits: Kimi's trillion-scale K-line and GLM as walkable peaks rivaling the sealed ones; the fits-by-total/streams-by-active reminder; the geography of who builds them.
So for most people, even these arrive the way closed ones do — through operators who host them — just without lock-in. Openness changed who can host, not who must. And the geography is the story. These frontier-scale open weights come predominantly from Chinese labs — Zhipu's GLM, Moonshot's Kimi, DeepSeek, Alibaba's Qwen. Read that as fact, not verdict: for downloadable frontier-scale models, the center of gravity has shifted. Anyone planning around open weights should know where the biggest ones now come from. Building on them? Run episode one-thirty-four's license ritual — the seals are mostly permissive, always worth reading. Face episode one-eighteen's hosting math at giant scale. And weigh the real draw: sovereignty. Weights you hold don't vanish when a provider changes terms or a border hardens.
The trap to avoid
Assuming frontier-scale means American and closed — the largest downloadable models invert both assumptions.
Why it matters — and what’s next
But the absolute ceiling is still sealed.
For some that outweighs everything; for others the hosting burden decides. Both read the same ledger honestly. The trap is a pair of stale defaults: frontier-scale means American, and frontier-scale means closed. The open trillion-summits are neither — foreign and open, the largest ownable brains on Earth. You don't need an opinion about that to need an accurate map of it: supply chains, procurement, and risk planning sit on top of it. Update the map. So open weights climbed to trillion scale — and yet the absolute ceiling still sits behind glass, across the valley, higher than the tallest walkable summit, with the newest capabilities appearing there first. Who holds that ceiling, what it buys that ownership can't, and what you surrender for it: next. The closed frontier — GPT-5, Fable, and Mythos.
This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.
Full transcript 2:45 of narration
Last episode's open ridge grew summits nobody predicted: trillion-parameter models you can download, own, and run — rivaling the sealed peaks in scale. And their provenance rewrites the map. The biggest brains anyone can legally own come, overwhelmingly, from labs the early family tree barely mentioned. The open trillion era.
The summits: the GLM line, Kimi's K-series, the DeepSeek giants — all mixture-of-experts, hundreds of billions to a full trillion total. And crucially, most ship under permissive seals — genuinely ownable, not just visible. A model with a trillion parameters and an open license would have sounded like a contradiction three years ago. It's now a category with several members.
But recall episode one-twenty-three, because marketing won't. These fit by total, stream by active: a trillion-parameter open giant still demands rack-scale memory to host. Ownable is not laptop-runnable. So for most people, even these arrive the way closed ones do — through operators who host them — just without lock-in. Openness changed who can host, not who must.
And the geography is the story. These frontier-scale open weights come predominantly from Chinese labs — Zhipu's GLM, Moonshot's Kimi, DeepSeek, Alibaba's Qwen. Read that as fact, not verdict: for downloadable frontier-scale models, the center of gravity has shifted. Anyone planning around open weights should know where the biggest ones now come from.
Building on them? Run episode one-thirty-four's license ritual — the seals are mostly permissive, always worth reading. Face episode one-eighteen's hosting math at giant scale. And weigh the real draw: sovereignty. Weights you hold don't vanish when a provider changes terms or a border hardens. For some that outweighs everything; for others the hosting burden decides. Both read the same ledger honestly.
The trap is a pair of stale defaults: frontier-scale means American, and frontier-scale means closed. The open trillion-summits are neither — foreign and open, the largest ownable brains on Earth. You don't need an opinion about that to need an accurate map of it: supply chains, procurement, and risk planning sit on top of it. Update the map.
So open weights climbed to trillion scale — and yet the absolute ceiling still sits behind glass, across the valley, higher than the tallest walkable summit, with the newest capabilities appearing there first. Who holds that ceiling, what it buys that ownership can't, and what you surrender for it: next. The closed frontier — GPT-5, Fable, and Mythos.