aiqbench takeAI Products

DeepSeek's V4.1 Flash-1017 release centers on token efficiency rather than new capabilities, with the company claiming over 50% fewer tokens in most scenarios while keeping API pricing unchanged. These figures are self-reported, and the cost comparisons against unnamed rival models lack independent benchmarks, so readers evaluating total cost of ownership should run their own workloads before drawing conclusions.

Sources covering36Kr, Fortune, Interconnects, Latent Space

What this isA model's reading of the article and related coverage, written for aiqbench. Not the publisher's words.

DeepSeek新模型Token消耗降超五成

36KrOctober 18, 20261 min read

DeepSeek正式推出大模型新版本V4.1 Flash-1017。在国产人工智能大模型普遍处于迭代关键期的当下，该版本率先完成升级，整体性能实现显著跃升。此前已有公开数据显示其技术指标，此次DeepSeek更首次自主发布帕累托前沿曲线，直观呈现模型在效率与成本维度的综合表现。V4.1 Flash-1017在此曲线中达到行业领先水平：单位任务综合成本仅为Fable、Astra等同类模型的百分之一至二十分之一，性价比优势达四五十倍以上。本次发布的1017版本虽维持原有API价格体系，但在Token使用效率方面完成深度重构——通过任务感知型压缩、动态上下文裁剪及自适应解码策略，显著降低各类型请求的实际Token消耗。实测表明，多数场景下Token用量较前版下降逾50%，相应TCO水平亦将同步迎来新一轮实质性下探。（新浪财经）

Read the full article at 36Kr

Read at 36Kr ↗Ask about this article