A Beijing lab just released the largest open-weight AI model in the world, and it did so by publishing the blueprint instead of guarding it.
The Bottom Line
- Kimi K3 is the largest open-weight AI model released to date at 2.8 trillion parameters, roughly 75 percent bigger than DeepSeek’s V4 Pro, giving developers a frontier-scale system they can inspect and fine-tune rather than rent through a closed API.
- Independent evaluators including Arena.ai, Vals AI, and Artificial Analysis rank Kimi K3 ahead of Claude Opus 4.8 and GPT-5.5 on coding and agentic tasks, though it still trails Anthropic’s Fable 5 and OpenAI’s GPT-5.6 Sol on overall performance.
- Full model weights are not out yet. Moonshot has committed to publishing them on July 27, 2026, under a modified MIT license. Until then, Kimi K3 is only usable through Moonshot’s API and apps.
- The release triggered a sharp same-day selloff in shares of rival Chinese AI firms Zhipu (Z.ai) and MiniMax, a sign that China’s crowded AI field is starting to consolidate around fewer, better-capitalized leaders.
- Moonshot is closing a funding round that could value the company above 30 billion dollars and has begun preparing a Hong Kong listing, meaning Kimi K3 is as much a fundraising and IPO story as a technical one.
For founders and IT decision-makers weighing AI vendors, an open-weight model at this scale changes the conversation around lock-in, even if running it in-house remains out of reach for most companies.
Kimi K3 Enters as the Largest Open-Weight Model Yet
Moonshot AI, a Beijing startup founded in 2023, released Kimi K3 on July 16. At 2.8 trillion total parameters, the company describes it as the first open model to approach the 3-trillion-parameter mark and the largest open-weight system released anywhere. That puts it well ahead of the two other Chinese models that had held the open-weight lead this year, DeepSeek’s V4 Pro at 1.6 trillion parameters and Zhipu’s GLM-5.2 at 744 billion.
The release lands at a moment when the assumption that Chinese AI labs trail US frontier developers by months has been repeatedly tested. Z.ai’s GLM-5.2 scored close to top American closed-source models earlier this year, and DeepSeek’s releases have already reshaped how much compute a competitive model is assumed to require. Kimi K3 pushes that pattern further by combining frontier-adjacent benchmark scores with a release strategy that hands the model itself to anyone who wants to build on it, rather than keeping it behind an API wall.
One framing worth correcting for context: some coverage tied to this launch has referenced Anthropic’s Fable and Mythos models as having been withdrawn over security concerns. The actual sequence is more specific. Fable and Mythos launched June 9, 2026. Anthropic suspended access on June 12 to comply with US Department of Commerce export controls, not a security recall. The Commerce Department lifted those controls on June 30, and Anthropic restored access on July 1, two weeks before Kimi K3 shipped.
Architecture Built for Efficiency at Scale
Kimi K3 does not fire all 2.8 trillion parameters on every request. It runs on a sparse mixture-of-experts design that activates just 16 of its 896 expert sub-networks per query, or roughly 1.8 percent of the total model. That routing is what keeps a model this large computationally workable at all.
Moonshot paired that structure with two architectural additions developed internally: Kimi Delta Attention, a hybrid attention mechanism built for long conversations, and Attention Residuals, which the company says improves scaling consistency. Both had already been published as open research on Moonshot’s GitHub before this release. The model also carries a 1-million-token context window and native support for images and video, positioning it for long-horizon coding and agent-style workloads rather than single-turn chat.
How Kimi K3 Performs Against US Frontier Models
Because Anthropic and OpenAI do not disclose parameter counts for systems like Fable, Mythos, or GPT-5.5, a direct size comparison with Kimi K3 is not possible. Performance benchmarks fill some of that gap. Arena.ai ranked Kimi K3 first in a test measuring web interface-building ability, ahead of Fable 5. Vals AI placed it second overall, behind Fable 5 and ahead of GPT-5.6 Sol. Artificial Analysis found its results comparable to GPT-5.5 and Claude Opus 4.8 on complex, multi-step tasks in particular.
Omdia chief analyst Lian Jye Su has pointed to cost as the more decisive factor behind the model’s traction, noting that Chinese models can run at a fraction of what leading US systems charge, while cautioning that scale alone does not guarantee Kimi K3 is the best performer for every task. That caution matters here, because the full weights will not be public until July 27. Everything currently known about Kimi K3’s real-world behavior comes from Moonshot’s own API and app releases, not from independent parties running the model themselves.
Behind Moonshot: A Fast Climb and an IPO in Motion
Moonshot was founded in March 2023 by Yang Zhilin, Zhou Xinyu, and Wu Yuxin, three Tsinghua University classmates. Yang, who holds a doctorate in machine learning and worked at Carnegie Mellon, Google Brain, and Meta, has led the company from a roughly 4.3 billion dollar valuation at the end of 2025 to about 20 billion dollars after a Meituan-led raise in May 2026. Moonshot is now closing a further round that could value it above 30 billion dollars, with backers including Alibaba, Tencent, Meituan, HongShan, IDG Capital, and China Mobile.
Kimi K3’s release also coincided with the company circulating a shareholder resolution to prepare a Hong Kong listing within roughly six months. The stock market reaction to the launch was immediate on the Chinese side as well. Shares of domestic rivals Zhipu and MiniMax fell 27.7 percent and 16.5 percent respectively in Hong Kong trading the same day, a reaction that says as much about investor expectations for consolidation among China’s AI labs as it does about Kimi K3 itself.
What This Means for Businesses Evaluating AI Vendors
For a founder or IT lead comparing AI providers, Kimi K3 is not something most organizations will run themselves. Fellow at the American Enterprise Institute Ryan Fedasiuk has estimated that hosting a 2.8 trillion-parameter model locally would require hundreds of thousands of dollars in computing hardware, well outside what a mid-sized business would deploy on its own. What it does change is the reference point. A frontier-class open-weight model narrows the gap between what a company can build in-house and what it has historically had to rent from a small handful of closed-source providers. Whether that translates into lower costs or faster deployment for a specific business depends on the workload, and on how the July 27 weights release actually performs once independent developers get their hands on it.
Kimi K3 is available now through the Kimi app on iOS and Android, the kimi.com website, and the Kimi Work desktop app, with a free tier that carries usage limits. Full open weights are scheduled for July 27, 2026.
Monitor the systems, tools, and digital infrastructure decisions redefining competitive advantage in the Tech section of Hemos PH.
