Back
Moonshot AI unveils Kimi K3, a 2.8-trillion-parameter open-source model

Moonshot AI unveils Kimi K3, a 2.8-trillion-parameter open-source model

Moonshot AI showcased Kimi K3 at NVIDIA’s 2026 GTC after a $2 billion funding round, highlighting its open-weight design, multimodal features and low API pricing as it pushes into the global AI race.

Fact Check
The substantive claims are strongly corroborated by multiple independent top-tier outlets. Reuters, Forbes, and CNBC all confirm Moonshot AI released/unveiled Kimi K3 around July 17, 2026 as a 2.8-trillion-parameter open-weight model, described as the world's largest, competing with OpenAI and Anthropic, with multimodal capability and a 1M-token context window. CNBC and Forbes confirm the ~$2 billion funding round at a $20B+ valuation. Cryptobriefing confirms low API pricing (~$3/M input, ~$15/M output) and open weights. The one element with slightly weaker direct confirmation is the specific framing of the reveal as happening 'at NVIDIA's 2026 GTC': mainstream reports describe it as a product launch/release on July 17, 2026, while cryptobriefing and social evidence indicate Moonshot's founder Yang Zhilin was at GTC 2026 that same week and presented there. GTC 2026 coverage (Constellation Research) also references the Kimi K3 release in the GTC context, making the venue framing plausible though not the central verified fact. Because every core factual claim (2.8T params, open-source, multimodal, low pricing, $2B round, GTC-week timing) checks out, the claim is likely true.
Summary

Moonshot AI has showcased Kimi K3, a multimodal AI model with 2.8 trillion parameters and a 1 million-token context window, positioning it as a direct challenger to leading U.S. systems. The model, developed by the Chinese startup, is described as open-weight, allowing developers to download and run it themselves rather than rely only on paid API access, and it also accepts images alongside text for vision-based tasks. Moonshot has priced API access at $3 per million input tokens and $15 per million output tokens, extending a low-cost strategy that has already drawn attention in China’s AI market. The presentation came at NVIDIA’s GPU Technology Conference in 2026, shortly after Moonshot closed a $2 billion funding round in May 2026 that pushed its valuation above $20 billion. Yang Zhilin, who co-founded Moonshot AI in March 2023 and holds a PhD from Carnegie Mellon University, used the high-profile GTC stage to underline how quickly the company has scaled.

Terms & Concepts
  • open-weight: A model release approach that lets developers download and run the model themselves instead of relying solely on hosted API access.
  • multimodal: Able to process more than one kind of input, such as text and images.
  • parameters: Numerical weights in an AI model that are commonly used as a rough measure of model capacity.