Moonshot AI has released Kimi K3, a 2.8-trillion-parameter multimodal open-source model. According to Dean Ball, OpenAI's head of strategic futures, its performance in agent programming scenarios now matches the best publicly available models.
The model uses a mixture-of-experts architecture, activating only 16 of its 896 experts for any given task. It features a 1-million-token context window and native vision capabilities.
On key benchmarks, K3 scored 67.5 on DeepSWE and a 77.8 raw pass rate on ProgramBench. While it generally trails proprietary leaders like Claude Fable 5 and GPT-5.6 Sol in overall evaluations, it is competitive in sustained software engineering and agentic coding tasks.
Moonshot AI is a traditional AI laboratory. The model has no crypto or blockchain component. Its API pricing, at roughly $3 per million input tokens, sets a competitive benchmark that impacts the margins for decentralized compute providers.
The model's scale, requiring significant computational resources, reinforces an advantage for centralized providers. Its development also suggests that U.S. chip export controls have had limited effect on Chinese AI capability deployment.