The race to build the biggest, most capable open artificial-intelligence models just moved to a new scale. Moonshot AI, a Beijing-based laboratory, has announced Kimi K3 — a model with about 2.8 trillion total parameters, making it the largest open-weight model released to date. The weights are scheduled to go fully open by July 27, though the model is already available through the Kimi app, an API and a coding product called Kimi Code.
A parameter is essentially a number a model adjusts during training to shape how it responds to prompts. More parameters generally give a model more "memory" and expressive range, but they also make it more expensive to run. Kimi K3 tackles this trade-off with a "mixture-of-experts" (MoE) architecture: the full 2.8-trillion-parameter brain is split into many specialist sub-networks, and only a small active subset is turned on for any given question. The result is frontier-class capability without paying to run every parameter for every request.
Kimi K3 also ships with a one-million-token context window — roughly a thousand times the length of a typical chat exchange — and the company places it near the top on several reasoning, coding and agentic benchmarks, alongside the leading U.S. systems. Developers have also shown it producing high-quality 3D visuals and handling long-horizon tasks.
Why it matters goes beyond numbers. An open, frontier-level model from China narrows the gap with American companies and makes cutting-edge AI far easier to run, fine-tune and deploy without paying per-token fees. For researchers and small teams, that lowers the bar to the ground; for the industry, it signals that the next generation of AI capability will not sit inside one company's server room.
Knowledge takeaway: Kimi K3 combines 2.8 trillion parameters in a mixture-of-experts design, a one-million-token context window and open weights releasing July 2026; it benchmarks close to leading U.S. frontier models and shows how open, large-scale AI can become affordable and widely deployable.