2026-09-11 · Fuente：SiliconANGLE
The Chinese startup says multi-party tests put its open-weight V4.1-Flash ahead of DeepSeek-V4-Pro on performance, cost, speed and total runtime, with changes to V4-Pro requests starting Sept. 14.
DeepSeek — formally incorporated as Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co. Ltd. — introduced DeepSeek-V4.1-Flash on Thursday (Sept. 10), positioning the release as the smallest model in a newly unveiled architecture family. The Chinese artificial intelligence startup says the open-weight model outperforms its own much larger flagship, DeepSeek-V4-Pro, based on testing conducted by multiple parties.
According to the company, those evaluations put V4.1-Flash ahead of V4-Pro on four measures at once: performance, cost, speed and total runtime. The claim is notable given the size gap between the two systems — V4.1-Flash sits at the bottom of the new family, while V4-Pro remains DeepSeek's flagship offering. Casting a compact model as superior to a premium tier on cost and runtime, in addition to raw capability, is an unusual pitch in a market where larger models typically occupy the top slot.
The open-weight distribution is central to the announcement. By releasing weights that outside parties can run and evaluate, DeepSeek invites precisely the kind of multi-party testing it now cites as evidence that its smallest new model beats its biggest. For enterprises managing inference budgets, the emphasis on cost and total runtime alongside capability speaks directly to the economics of deploying large models at scale.
DeepSeek also set a transition marker for its flagship service: starting Sept. 14, requests sent to V4-Pro will be subject to changes detailed in the company's announcement. Buyers and developers will be watching whether the multi-party test results hold up in production, and how the smaller Flash model settles into a lineup that still includes its larger sibling.
Etiquetas: DeepSeek · DeepSeek-V4.1-Flash · DeepSeek-V4-Pro · Open-Weight Models · Large Language Models · Model Efficiency
萬安算交所 AIXX · AIXX · Exchange verificado · Ir al mercado