大模型
www.aibase.com ·
DeepSeek V4.1 Flash 发布：552B MoE 新架构，官方称全面超越 V4 Pro
据 AIbase 报道，DeepSeek 发布 V4.1 Flash，为 552B 参数 MoE 模型，采用新的 Causal-Encoder-Decoder 结构，输入激活仅 8B、输出激活 16B，并称 KV Cache 大幅压缩，HBM 需求降至上一代的 1/4。报道称该模型已同步至 DeepSeek API（模型名 deepseek-flash），旧版 V4 Flash 与 V4 Flash Vision Exp 停用；官方计划自 2026 年 9 月 14 日 12:00（北京时间）起将 deepseek-v4-pro 请求路由至 V4.1 Flash 并按新价格计费。腾讯 WorkBuddy、CodeBuddy 与 OpenCode 被列为官方合作伙伴并已集成，权重与技术报告称已上架 Hugging Face。上述性能对比与定价细节来自聚合报道，尚待官方一手信息核实。
AI推荐理由 内容涉及 DeepSeek 新模型 V4.1 Flash 的架构与定价细节：552B MoE、Causal-Encoder-Decoder 非对称结构、输入激活 8B/输出 16B、KV Cache 大幅压缩（HBM 需求降至 1/4）、API 已同步并计划于 9 月 14 日将 V4 Pro 请求路由至 V4.1 Flash、腾讯 WorkBuddy/CodeBuddy 与 OpenCode 集成、权重与技术报告上架 Hugging Face。这些对模型选型、Agent 成本与开发工具链有实际参考价值。但来源为 AIbase 聚合站，非官方一手发布，正文未给出可核验的基准数据与官方链接，'全面超越 V4 Pro'等结论缺乏证据支撑，需以官方公告为准。
78
正文
DeepSeek V4.1 Flash Officially Released: 552B MoE New Architecture, Fully Exceeds V4 Pro
AIbase基地
Published inAI News · 4 min read · Sep 10, 2026
31.0k
DeepSeek V4.1 Flash model is released. It is the smallest model in its new series of model structures, with native multi-modal visual understanding capabilities. The design of the new model structure aims to achieve higher capability limits, faster reasoning speed, greater throughput, and scalability to larger parameter models.
Input Activation Only 8B, HBM Requirement Reduced to 1/4
DeepSeek V4.1 Flash is a 552B parameter MoE model that adopts a new Causal-Encoder-Decoder structure, with asymmetric input and output—input activation is only 8B, while output activation is 16B, significantly lower in cost than similarly sized models known so far. The new model also uses a new pre-training method and has been further trained on a larger scale through reinforcement learning, successfully surpassing the intelligence level of several flagship models including DeepSeek V4 Pro in benchmark tests.
In addition, the new generation model greatly reduces the size of the KV Cache, reducing the requirement for HBM to 1/4 compared to the previous generation, and for SSD to 1/8; compared to the original model, the KV Cache is only 1/437. In Agent usage scenarios, cache hit costs often account for a large proportion, and the compression of KV Cache significantly reduces the usage cost of Agent-type tasks.
API Has Been Synchronized, V4 Pro Will Be Routed on September 14
DeepSeek V4.1 Flash has been synchronized on the DeepSeek API, natively supporting multi-modal, and can be called by changing the model name to deepseek-flash. The old version models V4 Flash and V4 Flash Vision Exp have been discontinued. For compatibility reasons, the model names deepseek-v4-flash and deepseek-v4-flash-vision-exp will be temporarily routed to V4.1 Flash.
After multiple tests, V4.1 Flash has comprehensively surpassed V4 Pro in performance, cost, speed, and total time indicators. Therefore, the official plans to orderly discontinue V4 Pro: after 12:00 Beijing Time on September 14, 2026, until the release of V4.1 Pro, requests accessing deepseek-v4-pro will be redirected to V4.1 Flash and billed at the V4.1 Flash price.
Tencent (WorkBuddy, CodeBuddy) and OpenCode, as official partners, have now fully integrated DeepSeek V4.1 Flash. Thanks to architectural innovations, the official has accordingly reduced the pricing of V4.1 Flash, while still adopting peak-valley pricing—off-peak prices are half of peak-hour prices. The new pricing takes effect starting at 12:00 on September 10, 2026. Model weights and technical reports have been synchronized and made available on Hugging Face.
This article is from AIbase Daily
Scan to view
Welcome to the [AI Daily] column! This is your daily guide to exploring the world of artificial intelligence. Every day, we present you with hot topics in the AI field, focusing on developers, helping you understand technical trends, and learning about innovative AI product applications.
—— Created by the AIbase Daily Team
© Copyright AIbase Base 2024, Click to View Source -