Kimi K3: Moonshot AI Releases Open-Weight Model with 2.8 Trillion Parameters
Published July 14, 2026
Moonshot AI has unveiled Kimi K3, its most powerful model yet and what the company calls the first open-weight model in the 3T class. The model packs around 2.8 trillion parameters and supports a context window of up to one million tokens.
Designed for long-running coding, scientific research, knowledge work and complex agent workflows, Kimi K3 processes text, images and video natively. Moonshot AI concedes it does not yet match the strongest proprietary models across the board, but reports competitive results in several coding, agent and productivity benchmarks.
A massive MoE that stays efficient
Despite its size, Kimi K3 is built as a Mixture-of-Experts (MoE) model with 896 specialized experts, of which only 16 are activated per token. A router selects the relevant experts for each input, so the full parameter set is never computed at once.
Three components underpin the architecture: Kimi Delta Attention (KDA) for efficient long-sequence processing, Attention Residuals for better information flow between layers, and Stable LatentMoE for routing. Together with new training and data methods, Moonshot AI reports roughly a 2.5x improvement in scaling efficiency over Kimi K2.
Long-running coding: the MiniTriton case
Kimis K3's coding abilities are the clearest selling point. In one trial the model was given up to 24 hours to analyze, rewrite and accelerate GPU kernels, matching the strongest tested models. A larger project, MiniTriton, had Kimi K3 build its own GPU programming system — a domain-specific language, an MLIR-based intermediate representation, optimization passes, PTX code generation and a runtime — achieving performance comparable to Triton and torch.compile.
In a separate 48-hour autonomous run, an early version produced a chip design for a small model based on its own architecture, using open EDA tools and the Nangate 45 nm library. The simulated design occupied under four square millimeters and reached a 100 MHz clock. These are vendor case studies, however, and not independently verified.
Benchmarks: strong, but not the overall leader
Moonshot AI's benchmark results are mixed. On Terminal-Bench 2.1 Kimi K3 scores 88.3, just behind GPT-5.6 Sol. It leads the shown comparison on Program Bench with 77.8 and on the long-horizon SWE Marathon with 42 points. The company notes that some models ran under different agent environments (Kimi Code, Claude Code or Codex) and that not all results come from identical test runs — so the numbers are indicative rather than fully neutral.
Heavy infrastructure and clear limits
Training and serving use quantization-aware training with MXFP4 weights and MXFP8 activations. Moonshot AI recommends "supernode" configurations with at least 64 accelerators for efficient self-hosting, making full on-premise deployment far from a typical local AI setup. For most users, access via Kimi.com, Kimi Work, Kimi Code or the API will be more realistic.
The company also lists concrete weaknesses: Kimi K3 is sensitive to incomplete conversation history, can become unstable when switching from another model mid-session, and tends to take excessive initiative on ambiguous tasks — developers should set clear boundaries via the system prompt or an AGENTS.md file.
Open weights under a custom license
The full model weights are now available through an official GitHub repository under a dedicated Kimi-K3 license. This makes Kimi K3 more precisely an open-weight model: the weights can be downloaded and self-hosted, but the specific license terms must be reviewed before use.
Assessment
The showcased experiments are impressive, but vendor-internal benchmarks, differing test environments and the enormous hardware requirement make direct comparisons and practical self-hosting difficult. Still, Kimi K3 is a meaningful step for the open-weight ecosystem: models of this scale were previously available almost exclusively through closed platforms. Researchers, companies and infrastructure providers now get access to an open multimodal 3T-class model — provided they have the hardware or rely on a specialized provider.
