Kimi K3: Moonshot ships a 2.8-trillion-parameter open model
Moonshot AI released Kimi K3 on 16 July 2026 — reported as the largest open-weight model yet at 2.8 trillion parameters, with a one-million-token context window.
Beijing-based Moonshot AI released Kimi K3 on 16 July 2026 and describes it as the world's first open 3T-class system — reported as the largest open-weight AI model released to date.
The specification, as reported
- 2.8 trillion total parameters, in a mixture-of-experts (MoE) architecture
- Activates 16 of 896 experts per token — roughly 1.8% of the expert pool on any given step
- 1,048,576-token context window (one million tokens)
- Multimodal input: text, images and video, moving the Kimi line past the text-first K2 generation
- Architecture notes cite KDA and attention residuals for efficiency
The sparse activation is the whole trick here. A 2.8T-parameter dense model would be unservable for almost anyone; activating under 2% of the experts per token is what makes a model this size a practical proposition at all.
How Moonshot positions it
Moonshot's own framing is unusually candid: the company says K3 still sits behind Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol on overall performance, while claiming it outperformed every other model in its evaluation suite — including Claude Opus 4.8 and GPT-5.5 — across coding and agentic benchmarks. Independent rankings have also placed it competitively against Claude Opus 4.8.
It is pitched at long-horizon agentic work: navigating large repositories, calling tools, debugging, and iterating against images, logs, tests and runtime feedback.
The catch on "open"
This is the part to read carefully. Moonshot called K3 an open model and promised full weights by 27 July 2026, but as of guides updated on 20 July the weights, the final licence and the technical report were not yet public. Until the licence text is published, nobody can say what commercial use is actually permitted — see our explainer on open weights versus open source for why that distinction matters more than the headline number.
What it takes to run
Nothing about a 2.8T-parameter model is casual. Even with MXFP4-style quantisation and sparse activation, this is datacentre-class hardware, not a workstation download. For most teams the realistic route to K3 is a hosted API endpoint; the open weights matter more as a signal — a Chinese lab publishing frontier-adjacent weights while working around US compute limits — than as something you will personally deploy.
