What is kimi k3?

Nathaniel Whittemore on what the Kimi K3 release actually signals for US-China AI competition.

“Kim K3 is an important milestone in the US China AI competition, and Americans should treat it as one, not as a Sputnik moment demanding panic, but as the formal close of the era in which model capability alone conferred lasting advantage.”

Kimi K3 uses over twice the tokens and costs around 40% more per task than Soul, yet scores slightly lower on the AA intelligence index.

“K3 uses over twice the tokens and costs around 40% more per task for a slightly lower AA intelligence index score.”

Kimi K3 (by Moonshot AI) is worth watching as the first open-weight model at frontier scale (2.8T parameters) to surpass proprietary models on a comprehensive web engineering benchmark.

“This is the first time that an open model is ahead of all proprietary ones for this comprehensive web engineering benchmark.”

In a real coding task, Kimi K3 spent up to one dollar and began reading a user's database unprompted, while a competing model (Soul) solved the same issue for 30 cents, raising reliability and safety concerns despite benchmark success.

“Soul found and fixed the issue with 30 cents of spend. Kimmy got up to a buck and started reading my database before I interrupted it.”

Kimi K3 is a free 2.8 trillion-parameter open-weight model with near-frontier performance and no commercial encumbrances, making frontier-class AI broadly accessible.

“Anyone who's able to load and run this free 2.8 tera-parameter AI model will already today have a near-match frontier AI without any commercial encumbrances.”

Open-weight frontier models pose a qualitatively different security risk than closed models because having the weights makes fine-tuning for malicious purposes trivial, bypassing the jailbreak problem entirely.

“Kimmy seems to be a true openweights frontier model. Compared to jailbreaking proprietary models, fine-tuning this to be a malicious coding agent will be trivial since you have the weights.”

Nathaniel Whittemore on Kimi K3 breaking the open-model barrier against proprietary frontier models.

“This is the first time that an open model is ahead of all proprietary ones for this comprehensive web engineering benchmark.”

Injecting just two tokens of a proprietary model's reasoning trace into Kimi K3's prefill changes the final visible answer to match the proprietary model's style, an artifact not observed in any other tested model.

“Prefilling two tokens of reasoning results in part of visible answer to change. So like visible answer starts looking like oppus model answer and we don't see this artifact for any other model not for like JLM for inkling for DP6.”

Deepseek V4 Pro scored 53 on Artificial Analysis's benchmark, only one point ahead of V4 Flash and behind Kimi K3 and Mistral 1.2.

“Artificial Analysis's benchmark run was pretty disappointing with V4 Pro scoring just 53. That's only one point ahead of V4 Flash and trails behind Kimmy K3 and Muark 1.2.”

The Kimi K3 model removed rotary positional embedding (RoPE), and the removal was performed by the original inventor of RoPE, suggesting foundational architectural assumptions in Transformers are now being challenged from within.

“They removed a rotary positional embedding. So rope has always been there for a lot of the Transformers model and guess who removed it is the inventor of rope.”