Shanghai-based AI startup StepFun released Step 5 Preview, a 600B-parameter sparse Mixture-of-Experts (MoE) model that activates just 27B parameters per token to deliver frontier coding and financial reasoning performance at $1 per million input tokens. Meanwhile, Alibaba Cloud expanded its open-weights lineup with Qwen-Image-2.1, a 7B unified architecture capable of simultaneous native generation and layer-based editing with RGBA transparent background support.

We’ll talk about:

• A 600B sparse MoE model featuring a 1M-token context window, scoring 67.7 on DeepSWE v1.1 and 66.4 on FrontierFinance at a fraction of frontier inference costs.
• A fast 7B parameter image model designed for unified generation and multi-reference editing with native RGBA transparency support.
• Autonomous agent startup ...