The largest open model ever built just shipped, and it did not come from a US lab.

On July 16, Moonshot AI, the Alibaba-backed Beijing startup behind the Kimi series, released Kimi K3: a 2.8-trillion-parameter model it calls the biggest open-source AI ever made. It landed the week of Shanghai's World AI Conference, it benchmarks neck-and-neck with the best closed models from Anthropic and OpenAI, and it costs a fraction of what they charge. After a full week of leaks, the real thing is finally here, and it is a bigger deal than the leaked promo page suggested.

The number that matters is 2.8 trillion. That is K3's parameter count, nearly triple Kimi K2.6's 1 trillion, and it makes K3 the largest open-weight model anyone has released, well past DeepSeek's 1.6-trillion V4-Pro and every other Chinese open model this year. Moonshot is calling it the first open 3-trillion-class model, its ninth open-source scale record in a single year.

It is built to compete at the top, not on price alone. K3 is a Mixture-of-Experts model with a 1-million-token context window priced flat across the whole window, new efficiency tricks Moonshot calls Kimi Delta Attention and Attention Residuals, and native vision input that produced the viral 3D generations flooding X on launch day. On the GDPval-AA real-world-task benchmark it lands third overall at 1,687, behind only Claude Fable 5 and GPT-5.6 Sol, and independent testing from Artificial Analysis put its intelligence score behind only Fable 5. It leads the field on a couple of coding benchmarks and trails GPT-5.6 Sol on Terminal-Bench by half a point.

The pricing is the twist. At $3 per million input tokens and $15 per million output, with cache hits down to $0.30, K3 costs roughly three to four times more than earlier Kimi models, the most a Chinese lab has ever charged. But it still undercuts Claude Opus 4.8 at $5 and $25 and GPT-5.5 at $5 and $30, so Moonshot is pricing K3 as a frontier flagship that happens to be cheaper than the West's, not as another budget play.

For a year the trade-off was simple: closed US models for the frontier, open Chinese models for the savings. K3 collapses that. A model this close to Fable 5 and Sol, that you will be able to self-host once the weights drop, changes the math for any team that wants frontier-adjacent quality without sending code to a US API or paying Fable's $10 and $50 rates. And the open weights are coming: Moonshot has promised them by July 27, the date self-hosting and regulated teams should circle. The bigger signal is momentum. Nine open-source scale records in a year, each one narrowing the gap, is the story US labs are actually worried about, and K3 is the loudest data point yet.

Read the fine print before you rip anything out. K3 ships with only one reasoning setting, max, which makes it slow and token-hungry: early testers clocked it around 28 tokens per second with multi-second first-token latency, and one simple test burned more than 13,000 reasoning tokens to produce a short answer. That is powerful but expensive per task in practice, not the cheap-and-fast Kimi you may remember. Most of the headline benchmarks are still Moonshot's own, with independent numbers only starting to land. The open weights are promised but not out until July 27. And the launch itself was quiet, a banner on the docs site rather than a full model card, so some details are still filling in. Treat it as a serious frontier option to evaluate, not a drop-in replacement to trust blind.

The stories moving fast beyond today's headlines:

Moonshot AI released Kimi K3 on July 16, a 2.8-trillion-parameter Mixture-of-Experts model it calls the largest open-source AI ever built, nearly triple Kimi K2.6 and well past DeepSeek's 1.6-trillion V4-Pro. It ships with a 1-million-token context window priced flat across the whole window, new efficiency tricks called Kimi Delta Attention and Attention Residuals, and native vision, and it benchmarks neck-and-neck with the top closed models: third overall on GDPval-AA behind only Fable 5 and GPT-5.6 Sol, with independent testing from Artificial Analysis placing it behind only Fable 5. Pricing is the twist at $3 and $15 per million tokens, three to four times pricier than earlier Kimi models but still under Opus 4.8 and GPT-5.5, so Moonshot is positioning K3 as a frontier flagship that happens to undercut the West rather than another budget play. The catch: it runs only one reasoning setting, max, which makes it slow and token-hungry, most headline benchmarks are still Moonshot's own, and the open weights are not out until July 27. The bigger picture is momentum, K3 is Moonshot's ninth open-source scale record in a year, and a self-hostable model this close to the frontier collapses the old trade-off between closed US quality and cheap Chinese open weights. For builders: try it on Kimi web, the API, or OpenRouter now, mark July 27 for the weights, and evaluate it as a serious option rather than trusting the self-reported numbers blind.

Stay building. 🤖