Alaya Token adapts Kimi K3: world's first open-source 3-trillion-parameter model goes live on Token Factory

Jiuzhang Yunji DataCanvas's Alaya Token, a professional token-scale delivery platform, has completed deep adaptation and production-grade deployment of the Kimi K3 model. This marks a key step in the company's 'AI Factory' strategy, following the integration of GLM-5.2 and DeepSeek-V4 Flash. Kimi K3 is the world's first open-source 3-trillion-parameter large model, using self-developed KDA mixed linear attention and attention residual architecture, with a 2.8-trillion-parameter MoE design supporting native vision understanding and a 1-million-token context window. It leads open-source models in programming, agent, and complex reasoning benchmarks, scoring 1679 on Frontend Code Arena and 76.8% on SWE-bench Verified. Alaya Token's core value lies in an 'aggregate but not lock' ecosystem, allowing enterprises to switch models with a single key. For Kimi K3, the engineering team optimized low-level operators and KV-cache dynamic scheduling for long-context inference. Alaya Token serves high-value ToB scenarios like financial risk control, industrial quality inspection, and code generation.