-+ 0.00%
-+ 0.00%
-+ 0.00%

Step 5 Preview weights were officially released today, according to Step News. This release: BF16 and FP8 model weights, technical blogs, supporting open source development tools. According to the introduction, Step 5 Preview is designed for long-range agent tasks. Using a sparse MoE architecture, the total number of parameters is 600B, each token activation parameter is only 27B, supports 1 million token context windows, and natively supports text and visual input. Currently, model weighting has launched Hugging Face and Model Scope, and provides adaptation support for various chip platforms and inference frameworks.

Zhitongcaijing·10/15/2026 03:17:05
Listen to the news
Step 5 Preview weights were officially released today, according to Step News. This release: BF16 and FP8 model weights, technical blogs, supporting open source development tools. According to the introduction, Step 5 Preview is designed for long-range agent tasks. Using a sparse MoE architecture, the total number of parameters is 600B, each token activation parameter is only 27B, supports 1 million token context windows, and natively supports text and visual input. Currently, model weighting has launched Hugging Face and Model Scope, and provides adaptation support for various chip platforms and inference frameworks.