What Moonshot launching Kimi K3 against OpenAI and Anthropic means — July 17, 2026

Radar Flash Edition. China's Moonshot AI has launched Kimi K3, a 2.8 trillion-parameter model presented as the world's largest open-weights model, positioning it directly against systems from OpenAI and Anthropic. It is a mixture-of-experts model that only activates a small fraction of its parameters for each response, featuring a one-million-token context window, native vision understanding, and a new attention mechanism (Kimi Delta Attention) that accelerates generation up to 6.3 times in long-context scenarios. It is already available via the Kimi app and API — compatible with the OpenAI SDK, at $3 per million input tokens and $15 per million output tokens —, and the weights will be released on July 27 under a modified MIT license. Source

Why it matters: according to benchmarks published by Moonshot itself, K3 competes with the best closed models from OpenAI and Anthropic in real-world task tests — "close to leading closed-source models," in the company's words —, although independent evaluations to confirm this are still pending. The relevant point isn't the parameter count, but the move: the open frontier is once again nipping at the heels of the closed frontier, just as DeepSeek did in January 2025. If anyone can host a model of this caliber, inference providers will offer it within weeks, competing on price, which puts pressure on rates across the board — including those of the closed models you may already be paying for.

For your company: three practical takeaways. First: more competition in the high-end segment means bargaining power — if you have a contract with OpenAI, Anthropic, Microsoft, or Google, you have a new argument for renegotiation, and testing K3 takes only an afternoon because its API is compatible with the OpenAI SDK. Second: data caution — using the Moonshot API sends your data to servers of a Chinese company, so do not input personal or confidential data without reviewing GDPR compliance; the sensible path for a European company will be to use the open weights via a provider with EU data residency once they are released on July 27. Third: don't migrate based on manufacturer benchmarks — they are marketing until an independent party replicates them; test with your own documents and processes, measure, and decide. And design your automations to be able to switch models without rebuilding the system: weeks like this are exactly why. Today's edition of the Daily Radar has the rest of the news.

Frequently Asked Questions

What is Kimi K3 and who launched it?

It is the new model from the Chinese startup Moonshot AI, launched on July 16, 2026: 2.8 trillion parameters in a mixture-of-experts architecture, one-million-token context, and vision understanding. It is available on the Kimi app and via API, and the weights will be published on July 27 under a modified MIT license, which would make it the largest open-weights model in the world.

Is it better than models from OpenAI and Anthropic?

It's not yet known: the benchmarks where K3 approaches or surpasses closed models were published by Moonshot itself, and independent evaluations are missing. The only valid comparison for your business is testing it with your own use cases and measuring quality, cost, and speed against what you are currently using.

Can a small or medium-sized company use Kimi K3 today?

Technically yes, via API and using the same OpenAI SDK you probably already use. However, data travels to Moonshot's servers, so do not send personal or confidential information without a prior GDPR analysis. Self-hosting a model of this size is not realistic for an SME; the practical option will arrive when European providers offer it following the release of the weights on July 27.