Kimi K3 goes live: Moonshot ships the model after the preview flood

In plain terms

Kimi K3 is available. After several days of videos and benchmark captures, Moonshot AI has made it accessible on its public interface (kimi.com) and via its developer documentation for the "code" mode. We move from the launch noise to the real test: does it work, at what price, and on which tasks does it outperform others.

The fact

On July 16, the pages kimi.com/en and kimi.com/code/docs/en/kimi-code/models confirm that Kimi K3 is in service. The Hacker News coverage at the top of the front page ("Kimi K3 Is Live") acknowledges the practical existence of the model. No new architectural details have been published during the live event compared to the preview phase; API pricing, context window served in production, and latencies are to be found on the product pages and cross-referenced with the independent measurements of the first users.

Under the hood

The preview had positioned K3 on agentic benchmarks and coding (see publi #1164). The transition to GA changes two things: (1) the API price moves from marketing promise to a comparable figure; (2) the real load tests the inference stack - a model that performs well in a demo does not necessarily hold up in production. The first third-party benchmarks on SWE-bench Verified, LiveCodeBench, and Aider will settle the debate on honesty vs. cherry-picking in the coming days.

What it changes

For teams building a coding agent: K3 joins the selection matrix (Claude, GPT, DeepSeek, Qwen), potentially at a Chinese price. For founders stacking LLMs behind a router, one more model means one more test, one more provider, and one more contract. Hosting outside China (Fireworks, DeepInfra, Together) becomes a key variable for Western companies under data residency constraints.

So what

Another launch in a saturated year is not an event in itself. But K3 arrives at a time when the pricing of coding agents is becoming the battleground (thread token-budget-caps). If Moonshot delivers performance close to the frontier at a Chinese price, the arbitrage shifts for tools that live on tokens. To watch: first third-party benchmarks, official pricing, presence with orchestrators (OpenRouter, LiteLLM).