ChinaDeepSeek’s model scores 77% at $0.20 a task, it says.
DeepSeek also open-sourced DeepInfer, the inference framework it says powers its production services.
DeepSeek reports shorter outputs and faster tasks, which can cut inference costs by using fewer tokens and less computing time.
Arena’s Agent Arena ranks it first among open-weight models but 11th overall, from 7,498 sessions.
DeepInfer runs on Nvidia SM100 and Huawei Ascend 950 chips, supporting only V4.1 Flash so far.
How each outlet framed it
- TechNode
- foregrounds inference cost per task ($0.11–$0.20) and flags the benchmarks as DeepSeek's own, unverified
Sources: TechNode