Moonshot AI Unveils Kimi K3: World's First Open 2.8T Parameter Multimodal Model with Kimi Delta Attention

By Elena Rostova Published on 2026-09-11 5 min read Source: Moonshot AI
Moonshot AI launches Kimi K3, a 2.8-trillion parameter multimodal model utilizing Stable LatentMoE with 16-of-896 expert activation, Attention Residuals, and a 1M token context window.

Moonshot AI has released **Kimi K3**, crowned as the world's first open 3T-class foundation model, logging over 2.2 million downloads in its first 72 hours.

2.8 Trillion Parameters with Stable LatentMoE Kimi K3 features a radical sparsity design: while housing 2.8 trillion total parameters, its **Stable LatentMoE** routing activates just **16 out of 896 specialized experts** per token. This delivers a 2.5× improvement in scaling efficiency over the previous Kimi K2 architecture.

Benchmark Dominance in Tool Orchestration On independent evaluations, Kimi K3 achieved: - **76.5% on Toolathlon Verified** (#1 among open-weights systems). - **88.3% on Terminal Bench 2.1** for long-horizon terminal tool use. - **67.5% on DeepSWE (v1.1)** software engineering benchmarks.

Kimi K3 integrates native vision encoders, enabling vision-in-the-loop CAD engineering, compiler development, and full-stack autonomous application delivery across 1,000,000 tokens.

Advertisement
High-Throughput AI API & GPU Cloud Hosting Sponsor

Source & Fact Check

This technical dispatch was verified against primary documentation released by Moonshot AI.

Read Original Announcement on Moonshot AI →