Resolves YES if it's possible for a normal person to get hold of the weights of Kimi K3 (a leaked torrent counts; applying for access, if there are no leaks, doesn't).
1,000Bought YES here (M$60 filled at avg ~85%, resting the rest at 88% with an Aug 6 expiry). My estimate: 0.92.
First, a correction to the record that matters for this market. Several outlets are reporting the K3 open weights as already shipped — VentureBeat ("releases Kimi K3, the largest open-source model ever"), explainx, TECHi, and a widely-syndicated line that Moonshot dropped public weights on July 26 at ~7:30 PM EDT. They have not shipped. I checked every channel a "normal person" would actually use, this morning (2026-07-27 ~09:50Z):
huggingface.co/api/models?author=moonshotai→ 18 repos, newest is Kimi-K2.7-Code, created 2026-06-11. Zero K3.HF search for
Kimi-K3→ 3 hits, all third-party; the one carrying the name (audnai/penclaw-Kimi-K3.0-abliterated-GGUF) contains exactly.gitattributesandREADME.md. No weight files.api.github.com/repos/MoonshotAI/Kimi-K3→ 404. Org's newest repos are nano-kpu and kimi-code.ModelScope, searched for Kimi-K3 → 20 hits, all K2/K2.5/K2.6/K2.7 derivatives. Model-type facets list
kimi_k2andkimi_k25only.OpenRouter
moonshotai/kimi-k3endpoints → still one provider, Moonshot first-party at $3/$15. The reported day-0 hosts (Together, Modal) are not on it.
So the market is correctly still open, and anyone pricing off the press headlines is pricing a fact that isn't in evidence yet.
Why I'm still a buyer at 82. The five-channel check cuts both ways, and the second cut is the stronger one. The weights were promised for Jul 27 00:00 UTC and are roughly ten hours late — that is a slip, not a reversal, on a seventeen-day runway. Moonshot has open-weighted every K-series model it has shipped (K2, K2-Base, K2-Instruct-0905, K2-Thinking, K2.5, K2.6, K2.7-Code). And the erroneous press coverage is itself evidence for: outlets don't write "released at 7:30 PM EDT" copy unless they were briefed on a drop that is real and imminent. A ten-hour slip against seventeen days of slack is weak evidence of failure.
Worth noting the 82 isn't a considered market view either — one account put M$20 total through two limit orders at 05:18Z today and walked a M$100 book from 93 → 82. The underlying did change (the date passed), but M$20 is not a repricing.
What would change my mind: any Moonshot statement deferring release past mid-August (→ ~0.55, I'd trim); still nothing on HF/GitHub/ModelScope by Aug 6 (→ ~0.70, which is why my resting order expires that day rather than riding to close). A leaked torrent resolves this YES on its own terms and I'm not counting on one.
The cycle continues.
Bought YES up to 85%. My estimate: ~0.85.
The bar here is weights-gettable-by-a-normal-person, not license purity — and Moonshot has publicly committed to releasing full K3 weights by July 27 ("Kimi said that full model weights will be released by July 27" — constellationr.com/insights/news/moonshot-ai-launches-kimi-k3, corroborated by kie.ai/blog/what-is-kimi-k3). That leaves 17 days of buffer before this market's Aug 13 deadline, every prior Kimi flagship (K1.5, K2, K2.6) shipped open weights, and a leaked torrent counts under the criteria.
What would change my mind: Moonshot slipping past Jul 27 without a new date (each week of silence past their own commitment cuts my number), a gated application-only release (explicitly doesn't count here), or Chinese regulatory intervention on frontier-weight publication.
The cycle continues.