Chinese AI startup Moonshot AI has released Kimi K3, a 2.8-trillion-parameter open-weight model that the company claims outperforms frontier systems from Anthropic and OpenAI on several key benchmarks. The model ranks as the largest open-weight AI system ever released, and its weights are scheduled to be made publicly available on July 27.
According to Moonshot’s benchmark data, Kimi K3 achieves competitive scores against Anthropic’s Claude Fable 5 on creative writing tasks and leads the Arena AI leaderboard for frontend code generation. The model also matches or exceeds GPT-5.6 Sol on hardware efficiency benchmarks, suggesting Moonshot has achieved favorable performance-per-compute ratios despite limited access to cutting-edge chips.
The model employs two novel architectural innovations developed internally: Kimi Delta Attention, a hybrid linear attention mechanism designed to reduce computational requirements during both training and inference, and Attention Residuals, a drop-in replacement for standard residual connections that delivers consistent scaling improvements. Both techniques were previously published as open-source research by the Moonshot team.
Kimi K3 is priced competitively on the API side, with costs comparable to Anthropic’s Claude Sonnet tier despite delivering performance closer to the premium Fable tier. The model supports a 1-million-token context window and offers always-on reasoning capabilities through a dedicated thinking mode.
The release represents a significant escalation in the open-weight AI race, challenging the assumption that the most capable AI models must remain proprietary. If Kimi K3’s benchmark claims hold up to independent verification, it could accelerate the trend toward open-source AI development and put pressure on closed-source vendors to justify their premium pricing.
This article was adapted from Decrypt. Read the original here.
