Search
DeepSeek made a quiet but significant move last night with the official release of its V3.1 version, which boasts an expanded context length of 128k. This iteration of the model impressively houses 685 billion parameters and offers support for a versatile range of precision formats, spanning from BF16 to FP8. Based on public insights and rigorous testing, V3.1 has shown marked improvements in code generation capabilities and frontend aesthetics. However, it still encounters challenges when tackling more complex tasks. In specific scenarios such as programming, its performance is on par with GPT-5 but trails behind Claude 4.0 and 4.1 in certain benchmarks. Despite these limitations, V3.1 underscores a steady trajectory of progress and continues to generate anticipation for future iterations.
-
C114 Communication Network -
Communication Home
