推文

@zzzXiao_guai · 2026-10-15 23:02

💡China's StepFun just open-sourced the full weights of Step 5 Preview — a 600B MoE built for long-horizon agent work. AA Intelligence Index: 44, top-3 among open models globally. The interesting part isn't the score. It's the price of the score. 🧵 2 The architecture is the story: 600B total parameters, only 27B active per token. 1M context, native text + vision. StepFun's bet: the next scaling law isn't more compute, it's more efficient compute. 3 Per Artificial Analysis, one Intelligence Index task costs ~$0.71 on Step 5 Preview. Claude Opus 5: $5.86 for the same task suite. Roughly 1/8 the cost at comparable intelligence. API pricing: $1 in / $2.7 out per 1M tokens. 4 And it's not a paper launch. On OpenRouter, Step 5 Preview is already #2 globally in token usage — daily and weekly charts, as of today. Open weights (BF16 + FP8) are live on Hugging Face and ModelScope. 5 The tech report has one detail worth stealing: context compression trained end-to-end, so the agent tidies its own task state mid-run instead of drowning in a 1M window. Open models aren't chasing the frontier anymore. On agent workloads, they're repricing it.

曝光 42 · 评论 0 · 点赞 0 · 书签 0