Today Moonshot AI dropped the full weights of Kimi K3 on HuggingFace. 2.8 trillion parameters, Modified MIT license, free to download, modify, and self-host.
It is the largest open-weight model humans have ever released. Full stop. But here is the number you should hold onto: on independent hallucination tests, it makes things up more than half the time.
You can download it tonight. Your hard drive cannot handle it.
The weights are roughly 594 GB. For scale: a 4K movie is a few GB. This is a hundred-plus movies. And downloading is the easy part. To actually serve it, you need about 1.4 TB of GPU memory, and Moonshot recommends pooling 64 accelerators into a single machine.
In plain terms: this is not something your studio runs. Sixty-four top-end cards. Small teams simply do not have that infrastructure. The only ones deploying day one are cloud providers, inference services, and a handful of giants.
2.8 trillion sounds terrifying. It only wakes 50 billion per token.
K3 is a mixture-of-experts model. The 2.8 trillion is the total; per token it activates only about 50 billion — 16 of its 896 expert layers. That is exactly how Moonshot can claim serving costs that compete with models a fraction of its nominal size.
Skill is skill, scale is scale. K3 is genuinely strong at coding: it tops the Frontend Code Arena at 1679, and ran a full chip design autonomously in 48 hours. But writing code and telling the truth are different cognitive jobs.
The real new finding: it hallucinates more than its predecessor
Independent testers ran K3 through the AA-Omniscience hallucination benchmark. It hallucinated on about 51% of cases. Its predecessor, K2.6, scored 39%.
Watch the direction. This is not progress; it is regression, and it regresses exactly where production use hurts most. Using it to write code or reason through a chain is fine. Using it as a knowledge base, a customer-service answer engine, or a generator of outward-facing reports is risky, because it will invent plausible nonsense you cannot see.
Not my scare tactic. Moonshot's own lead said K3's comprehensive ability "approaches" Fable 5 and GPT-5.6 but "still has gaps." The gap is factual reliability.
Data sovereignty: from today you can dodge the China API
At launch, the only way to use K3 was Moonshot's API — servers in China. For enterprises that means two things: data leaving your jurisdiction, and Chinese legal data obligations.
With weights open, teams with infrastructure can self-host. Data never leaves your room. Most API risk vanishes. The cost is those 64 cards.
One old bill to mention: in April 2026 Moonshot's infrastructure had a cross-user data breach the company never publicly addressed. Whether you count that in vendor risk is your call.
What this means for you
Bottom line: you probably cannot run K3 yourself. But as a buyer, your leverage just changed.
- Small teams, studios: forget local deployment. Use an inference provider (they run the weights for you); pricing will likely beat Moonshot's official API.
- Engineering teams: put K3 on code generation and review now. The coding lead is real, and an open frontier model entering the market pressures everyone's API pricing — walk into renewal talks with Anthropic or OpenAI with a harder spine.
- Knowledge-product builders: do not put K3 on outward-facing Q&A untreated. Add a verification layer. 51% hallucination is not disqualifying for everything, but it is disqualifying for some.
One line: the biggest brick in open-source history landed today. Good brick. Just do not use it as a load-bearing wall — especially not the wall labeled "everything it says must be true."