Moonshot AI has released Kimi K3, its flagship model for the open-weights era — alongside PerceptionBench, a new benchmark for multimodal perception.
The Open-Weights Race Heats Up
K3 arrives into the fiercest open-model competition yet. Moonshot’s release cadence through 2025–2026 — K2, K2 Thinking, K2.5, K2.6, and now K3 — has kept it locked in a sprint with DeepSeek’s V-series, Alibaba’s Qwen line, and Zhipu’s GLM family. Each lab ships frontier-adjacent weights within weeks of each other, and enterprise adoption increasingly flows to whoever’s open model leads agentic and coding benchmarks that month.
K3’s launch pairing with PerceptionBench signals where Moonshot sees the next differentiator: not just reasoning, but perception — how well models interpret the visual and contextual world agents operate in.
Read the announcement at kimi.com/blog/kimi-k3.