Kimi K3's open weights drop Monday: 2.8T MoE, 93% on SWE-bench
Moonshot's frontier MoE has been callable at $3/$15 for a week; the open weights land July 27. Top-3 on SWE-bench — but self-hosting needs a GPU cluster.
Copy markdown
Open weights land Monday on Hugging Face
Moonshot AI publishes Kimi K3's weights on July 27 under a permissive — but still unnamed — license, its first open release since K2. The Hugging Face page is already live for release alerts.
You can already call it at $3/$15
The API went live July 16 and is OpenAI-compatible — swap the base URL and go, on Moonshot's platform or OpenRouter. It's $3 in / $15 out per million tokens, roughly Opus-tier pricing for a top-3 coding model.
93.4% on SWE-bench Verified
Independent vals.ai scoring puts K3 third overall on SWE-bench Verified at 93.4%, with a 76.2 coding index a hair behind Claude Fable 5 and the #1 slot (1,679 Elo) on the Arena leaderboard.
2.8T params, ~50B active, 1M context
It's a Mixture-of-Experts transformer — 2.8 trillion total, 16 of 896 experts (~50B) live per token — with a 1M-token window, native vision, and a new linear-attention scheme (Kimi Delta Attention).
But you won't run it on your desk
Even in 4-bit MXFP4 the weights are ~1.4TB; Moonshot's floor is a 64-accelerator supernode (think 8 nodes of 8x80GB). A 512GB Mac Studio holds ~37%, a 5090 about 2% — plan to rent, not self-host.
The catch: weak on knowledge
K3 tops coding and agentic automation but trails badly on knowledge reliability (18 vs Fable's 40) — a strong agent and coder, a shakier general-purpose Q&A brain.