DeepSeek, a Chinese mainland AI startup, has launched its upgraded flagship model, DeepSeek-V3.1, promising faster performance and seamless support for domestic chips.
Featuring a hybrid inference structure, V3.1 lets users toggle between reasoning and rapid-response modes with a "deep thinking" button on both app and web platforms. Early adopters praise its advanced reasoning, especially in math and coding tasks, citing impressive results in breaking down complex problems and generating functional code – even building simple games from scratch.
Building on its 2024 breakthroughs, DeepSeek-V3.1 maintains cost-effectiveness while delivering top-tier accuracy. Its efficient architecture cuts operational costs, making the model an attractive alternative to some closed-source counterparts.
V3.1's UE8M0 FP8 precision format is fine-tuned for next-gen domestic chips, enabling faster computation and reduced memory use. Although specific chip partners remain undisclosed, this move underscores China's emerging semiconductor ecosystem and DeepSeek's commitment to local innovation.
With global developers seeking high performance without hefty price tags, DeepSeek-V3.1 could reshape AI adoption, offering a compelling blend of speed, capability, and affordability.
Reference(s):
DeepSeek drops upgraded V3.1 model optimized for Chinese-made chips
cgtn.com
