Short briefs on the AI industry, built from local-language reporting.
Why it matters
Bundling terminal hardware, models, an operating system, and monthly token quotas makes the physical device the distribution layer for StepFun's software, shifting agent monetization from cloud API fees to consumer hardware sales.
What we don't know
Whether StepX can manufacture and deliver the hardware at scale. No independent benchmark or teardown has verified device performance.
REA passes 110,000 GitHub stars linking models to decompilers
Utopai ranks second globally using post-trained MiniMax model
StartLux open-sources decision models outscoring Jev on benchmark index
AMD agrees to acquire World Labs for $8.2 billion
Honor ships Magic9 with Alibaba agent hitting 90% completion
AgiBot-backed Qiyuan prices home humanoids at about $3,000Our read on recent briefs, in one line each
StepFun's pitch is cost per task, not peak score: an average of $0.71 per Artificial Analysis task, which StepFun puts at one-eighth of Claude Opus 5's, for an open model near the top of the rankings.StepFun releases 600-billion-parameter Step 5 Preview model 
For DeepSeek, agent training now depends on CPU infrastructure as much as on GPUs: since V4.1, rollouts run on DSec, so preempting a GPU job no longer wipes them out.DeepSeek details system running 3 million daily agent sandboxes 
Training data gathered by teleoperating real robots is limited by how many robots and operators there are.MiFeng pays people to record chores for robot training 
