On July 16, 2026, Moonshot AI released Kimi K3, a model that immediately claimed the title of the largest open-weight AI system in existence. With 2.8 trillion total parameters — the first openly available model to cross the 3-trillion-class threshold — K3 is not merely an incremental step in model size. It represents a deliberate architectural bet on how to convert raw compute into usable intelligence at a scale where most organizations cannot even afford to experiment.
The model is built on a Mixture-of-Experts (MoE) architecture, meaning that although it holds 2.8 trillion total parameters, only a fraction — 16 out of 896 expert modules — are activated for any given input. This sparse activation pattern keeps inference costs far below what a dense model of equivalent size would demand, while still making the full parameter count available across the breadth of tasks the model handles. The architecture is further reinforced by two novel components: Kimi Delta Attention (KDA), which improves how information flows across long sequences, and Attention Residuals (AttnRes), which stabilize gradient propagation through the model's hundreds of layers. Together, these innovations yield roughly a 2.5× improvement in scaling efficiency over Moonshot's previous model, Kimi K2.
Kimi K3 supports a 1-million-token context window — enough to process entire codebases, lengthy legal contracts, or multiple long-form research papers in a single pass. It also has native vision capabilities, meaning it can interpret screenshots, diagrams, and photographs without relying on a separate vision encoder. Moonshot describes K3 as designed for "long-horizon" tasks: coding across large repositories, automated workflows, and knowledge work that requires sustained reasoning over tens of thousands of tokens of context.
The release strategy is notable for its staged openness. The model is available immediately on Kimi's web platform, Kimi Code for developers, and the Kimi API. Full model weights are scheduled for release by July 27, 2026 — a deliberate pace that allows Moonshot to coordinate with inference partners and open-source maintainers to ensure the model can actually be run outside Moonshot's own infrastructure. This matters because a model of this scale requires significant hardware: inference at 1-million-token context lengths is not feasible on consumer hardware as of mid-2026, though quantization and distillation approaches are expected to follow.
In benchmarks, Kimi K3 achieves frontier-level results across coding, mathematics, and reasoning evaluations. Moonshot reports that K3 consistently outperforms other open models and, while it still trails the most powerful proprietary systems — Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol — the gap has narrowed considerably. For context-sensitive tasks like long-context retrieval and multi-step code generation, K3 approaches or matches proprietary performance in several categories. The broader significance is geopolitical as well as technical: K3 marks the moment when a Chinese company — Moonshot AI, a Beijing-based startup — set the global frontier for open model scale, a position previously held by DeepSeek and, before that, by American labs.
Knowledge takeaway: Kimi K3 is Moonshot AI's 2.8-trillion-parameter open MoE model, the largest open-weight AI system released to date; it uses Kimi Delta Attention and Attention Residuals for a 2.5× scaling efficiency improvement; it supports 1M-token context and native vision; full weights will be publicly released by July 27, 2026; it represents a major milestone in open AI development from a Chinese startup, narrowing the gap with proprietary frontier models.