China's Moonshot AI Unveils Kimi K3, the Largest Open Model Yet

The News

Moonshot AI, the Beijing startup behind the Kimi chatbot, has released what it calls the largest open artificial-intelligence model ever built. Announced on 16 July, Kimi K3 carries 2.8 trillion total parameters, though its sparse mixture-of-experts design fires only about 50 billion of them per query, drawing on 16 of 896 expert networks. It handles a one-million-token context window and is natively multimodal.

The company will publish the weights on 27 July under a modified MIT licence, letting developers anywhere download and run the system for free. On Moonshot's own benchmarks, K3 ranks second overall, behind only Anthropic's Fable 5 and OpenAI's GPT-5.6 Sol. It scored 77.8 on Program Bench and 93.5 on GPQA-Diamond, and topped both American rivals on Arena.ai's blind front-end coding leaderboard, while lagging on harder tests such as FrontierSWE.

One caveat matters. No independent researcher has yet confirmed the parameter count or reproduced the scores, and Moonshot, reported to be chasing a $30 billion valuation, has every reason to lead with its best figures.

Why It Matters

The launch is a marker of how fast China has closed the gap on the American labs. News of it knocked AI and semiconductor stocks lower, a sign that markets now treat a Chinese open-weight model as a real competitive threat rather than a curiosity.

The strategic message is blunt. "China is not going to follow anyone on both AI technology and standards," George Chen of The Asia Group told MIT Technology Review, adding that the country intends to lead the world instead. That ambition was on display days earlier when Xi Jinping pitched China as an AI partner to developing nations at the World Artificial Intelligence Conference in Shanghai.

Open weights are the lever. When DeepSeek shook markets in early 2025 by matching frontier performance at a fraction of the training cost, it showed that giving models away can be a strategy, not a concession. Kimi K3 pushes that playbook to the very top of the size chart.

Indian Angle

For India, a free frontier-grade model is both a gift and a dilemma. The country still has no home-grown system at this scale. Startups such as Sarvam and Ola-backed Krutrim, plus the government's IndiaAI Mission, are building sovereign models, but their compute budgets are a rounding error next to Moonshot's. A downloadable 2.8-trillion-parameter model that any Bengaluru team can fine-tune shortcuts years of that effort.

The cost arithmetic is the sharpest angle. Indian developers pay in dollars for every API call to GPT-5.6 or Fable 5, an expense that stings when revenue arrives in rupees. A self-hosted open model converts that variable bill into a fixed infrastructure cost, and for fintechs like Razorpay or support teams running millions of queries, the savings compound quickly.

The catch is trust. A Chinese-origin model invites scrutiny from MeitY and enterprise security teams, even though open weights run entirely on local hardware and send no data abroad. Expect Indian CIOs to weigh that nuance, and expect fresh questions about why India still consumes other nations' models rather than shipping its own.

FAQ

When will Kimi K3 be available to download?

Moonshot AI plans to publish the weights on 27 July 2026, about ten days after the 16 July announcement. Until then, the headline benchmark figures rest on the company's own testing, which outside researchers have not yet verified or reproduced.

How does Kimi K3 compare with American models?

On Moonshot's benchmarks it ranks second overall, behind Anthropic's Fable 5 and OpenAI's GPT-5.6 Sol. It leads both on Arena.ai's blind front-end coding leaderboard but trails on tougher software-engineering tasks. At 2.8 trillion parameters, it is billed as the largest open model yet released.

What does this mean for Indian AI startups?

Open weights let Indian teams fine-tune a frontier-class model locally instead of paying dollar-denominated API fees. That lowers costs for firms such as Sarvam and Krutrim, but it sharpens the question of whether India can build, rather than merely adapt, models of its own.

Where can I read the original coverage?

MIT Technology Review covered the launch in its daily Download newsletter, citing Reuters' report that first described Kimi K3 as the world's largest open model. The full item is linked below.

This story was reported by MIT Technology Review. Read the full original coverage at MIT Technology Review.