
9月19日消息,DeepSeek 4.1 Flash 在速度与成本上实现显著跃升,将「强力 AI 跑在手持设备上」的行业愿景又推近一步。实测显示,该模型在多项传统基准上超越上一代,并能稳定击败 DeepSeek 4.0 Pro。
效率提升最直观的体现是本地部署成本:运行一套 DeepSeek 4.0 Pro 本地系统约需30万美元,而 DeepSeek 4.1 Flash 的部署成本约为其四分之一。性能飞跃的关键是一项名为 CSA2 的技术实现——它在模型层与层之间引入共享内存。
在标准 AI 架构中,每一层通常维护自己专属的键值(KV)记忆。DeepSeek 4.1 Flash 改用编码器-解码器结构:编码器创建一份全局共享记忆,解码器从这份共享记忆中读取,使信息在层间的访问效率大幅提升。官方数据显示,该架构将 KV 缓存占用降低了75%。这一结构还原生支持视觉理解——给模型一张经典游戏菜单截图,它能直接写出复刻这款游戏的代码。
不过该模型有一个明显代价:在给出答案前会「思考很久」,推理过程消耗的 token 量远高于常规模型。虽然单 token 价格仍然低廉,但思考阶段的 token 总量相当可观。对本地部署用户而言高 token 消耗近乎免费,但通过 API 或 Lambda 等服务使用的用户会注意到,延长思考阶段的 token 开销不容忽视。
此前 DeepSeek 官方介绍称,V4.1-Flash 是总参数552B 的多模态混合专家(MoE)模型,采用因果编码器-解码器架构,输入激活8B、输出激活16B参数,支持100万 token 上下文窗口;自9月14日起,V4-Pro 的 API 请求已全部路由至 V4.1-Flash。
本文来源:https://axbrief.com/en/column/claude-code-gpt6-astra-and-gemini-4-pro-push-new-workflow-and-benchmark-260919
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
On September 19, DeepSeek 4.1 Flash demonstrated a significant leap in speed and cost efficiency, bringing the industry vision of running powerful AI on handheld devices one step closer. Benchmark tests show the model surpassing its predecessor in multiple areas and consistently outperforming DeepSeek 4.0 Pro.
The most tangible manifestation of this efficiency gain is in local deployment costs: while a full-scale DeepSeek 4.0 Pro system requires approximately $300,000, the deployment cost for DeepSeek 4.1 Flash is only about a quarter of that. The key to this performance boost is a technical innovation called CSA2, which introduces shared memory between model layers to significantly enhance data access efficiency.
In standard AI architectures, each layer maintains its own dedicated key-value (KV) memory. DeepSeek 4.1 Flash revolutionizes this by adopting an encoder-decoder structure where the encoder creates a global shared memory, and the decoder reads from it — effectively sharing information across layers rather than duplicating it. Official data shows this reduces KV cache memory consumption by 75%. Additionally, the model natively supports visual-language understanding — given a game menu screenshot, it can write the game’s code directly.
However, there is a noticeable trade-off: the model takes significantly longer to “think” before outputting an answer, and its inference process consumes far more tokens than regular models. While the price per token remains low, the cumulative cost of the extended thinking phase can be substantial for users relying on API or cloud services. For local deployers, high token consumption may be less of an issue due to lower operational costs.
Earlier, DeepSeek officially described V4.1-Flash as a 552B-parameter multimodal Mixture of Experts (MoE) model based on a causal encoder-decoder architecture. It features 8B input and 16B output active parameters, and supports a context window of up to 1 million tokens. Since September 14, all API requests for V4-Pro have been routed to V4.1-Flash.