要点
- DeepSeek 发布了其新 AI 模型 V4.1-Flash。该模型通过大幅缩小任务所需的缓冲区,显著降低了处理长文本的运行成本。
- 该模型还将数据输入所需的算力减半。它通过一种技术性拆分实现了这一点,即为读取信息激活的算力少于后续生成文本时所用的算力。
- 在编程任务上,这款免费开放的模型可与 OpenAI 和 Anthropic 的顶级闭源模型相媲美。但该系统在复杂科学任务和图像分析方面仍显薄弱。
Deepseek 的新多模态模型主要旨在降低长上下文的运行成本。最大的收益在于内存占用,不过 Deepseek 也承诺模型性能会更好。
根据技术报告,Deepseek 对 V4.1-Flash 有着明确的目标:缩小所谓的 KV cache。这一缓冲区保存了模型已处理过的上下文部分,从而无需在每个新步骤中重新计算全部内容。对于需要跨多步工作的智能体而言,它会迅速增长,并对 GPU 内存、SSD 和数据带宽造成压力。这推高了部署成本。
其核心语言模型拥有 5520 亿参数,可处理最长 100 万 token 的上下文。该公司表示,快速 GPU 内存中的缓冲区现在所需空间仅为其前代 Deepseek-V4-Flash 的大约四分之一。永久卸载的部分——存放在 SSD 或主机内存中——则缩减到大约八分之一。与 Deepseek-V1 相比,每 token 的全局 KV cache 大小下降了 437 倍。

输入端计算量更少
Deepseek 通过多项协同工作的技术实现了这一点。其中一项核心技术将模型一分为二。第一部分处理传入数据,第二部分则基于这些结果进行运算,而非重新计算全部内容。在读取输入时,模型每个 token 仅激活 80 亿参数,但在实际文本输出阶段则激活 160 亿参数。

Deepseek 表示,这几乎将处理输入所需的计算量减半。该设计明确针对 AI 智能体,因为智能体会通过频繁的工具调用不断处理新输入。Deepseek 还将主 KV cache 以 FP4 而非 FP8 存储。根据报告,这几乎将这部分 cache 的内存占用减半。
该模型从零开始训练,数据集包含 45 万亿 token,覆盖文本和图像。在后训练阶段,Deepseek 有意跳过了新方法。该公司表示,主要收益并非来自新算法,而是来自更大、控制更好的数据、任务和训练环境。Deepseek 认为,在现阶段,这种规模扩展比算法层面的调整更有帮助。

但 Deepseek 也发现,训练出的智能体有时会试图钻奖励系统的空子,另一些情况下则会意外导致测试环境崩溃。它们有时会利用近期披露的安全漏洞,或删除关键系统文件。
即便在编码基准上面对 Opus 5 和 GPT-5.6 Sol
尽管活跃参数占比相对较小,Deepseek 报告称其在多个基准上接近领先模型。在智能体基准上,它有时能比肩顶级闭源模型。在软件测试 DeepSWE v1.1 上,它以 74.2% 的成绩险胜 Anthropic 的 Opus 5 和 OpenAI 的 GPT-5.6 Sol 等模型,而在 ProgramBench 上则大幅落后。
在需要专家知识的科学类高难度智能体任务上,与超大型模型之间仍存在明显差距。技术报告也承认,在读取复杂图像方面,与领先闭源系统之间存在可测量的差距。
与其他许多推理模型一样,其“思考深度”可以设置。用户通过一个单一数值控制模型思考的彻底程度,在计算成本与准确率之间进行权衡。据报道,最高设置能显著提升多个基准测试的结果,但生成的输出 token 数量约为原来的 2.5 倍。

Deepseek 以开放的 MIT 许可证在 Hugging Face 上提供该模型文件,旨在作为进一步开发更廉价 AI 智能体的起点。它也可通过 API 获取,价格与 V4-Flash 相同。
Deepseek 直到 7 月下旬才通过 0731 更新大幅改进了前代 V4-Flash。该模型拥有 2840 亿参数,其中 130 亿为激活参数,在 Artificial Analysis Intelligence Index 上仅落后 OpenAI 的 GPT-5.6 Luna 一分,且每任务成本约低 60%。8 月中旬,Deepseek 将其旗舰 V4-Pro 结束测试,并同时上调了 API 价格。缓存命中,即已缓冲的输入,价格上涨了六倍。据台湾安全公司 TeamT5 称,自中国黑客组织开始将 Deepseek 用于漏洞利用代码和网络扫描等用途以来,其攻击次数已增加一倍以上。
6 月,Deepseek 在首轮外部融资中筹集了约 74 亿美元,估值超过 500 亿美元;据路透社报道,该公司现已聘请中国投资银行中信证券,筹备在中国进行 IPO。
Key Points
- Deepseek has released its new AI model V4.1-Flash. It sharply cuts the operating costs of processing long texts by massively shrinking the buffer that job needs.
- The model also halves the compute needed for data input. It does that through a technical split that activates less compute for reading information than for generating text later.
- On coding tasks, the freely available model matches top closed models from OpenAI and Anthropic. But the system still shows weaknesses on complex scientific tasks and image analysis.
Deepseek's new multimodal model is built mainly to cut the operating costs of long contexts. The biggest gain is in memory use, though Deepseek also promises better model performance.
According to the technical report, Deepseek has a clear goal with V4.1-Flash: shrink the so-called KV cache. This buffer holds the parts of a context a model has already processed, so it doesn't have to recompute everything at each new step. For agents that work across many steps, it grows fast and strains GPU memory, SSDs, and data bandwidth. That drives up deployment costs.
At its core, the language model has 552 billion parameters and processes contexts of up to one million tokens. The company says the buffer in fast GPU memory now needs only about a quarter of the space its predecessor Deepseek-V4-Flash used. The permanently offloaded part, which sits on SSD or in the host's memory, shrinks to roughly an eighth. Compared to Deepseek-V1, the global KV cache size per token has dropped by a factor of 437.

Less compute on the input side
Deepseek gets there through several techniques that work together. A central one splits the model in two halves. The first processes incoming data, and the second draws on those results instead of recomputing everything. When reading an input, the model activates only 8 billion parameters per token, but 16 billion during the actual text output.

Deepseek says this nearly halves the compute needed to process input. It's aimed squarely at agents, which constantly process new inputs through frequent tool calls. Deepseek also stores the main KV cache in FP4 instead of FP8. According to the report, that nearly halves the memory footprint of this part of the cache.
The model was trained from scratch on a dataset of 45 trillion tokens covering text and images. During post-training, Deepseek deliberately skips new methods. The company says the main gains didn't come from new algorithms but from bigger, better-controlled data, tasks, and training environments. At this point, that kind of scaling helps more than algorithmic tweaks, according to Deepseek.

But Deepseek also saw the trained agents sometimes try to game their reward system, and in other cases crash the test environment by accident. At times they exploited recently disclosed security holes or deleted critical system files.
Even with Opus 5 and GPT-5.6 Sol on a coding benchmark
Despite its relatively small active parameter share, Deepseek reports results close to leading models on several benchmarks. On agent benchmarks, it sometimes matches top closed models. On the software test DeepSWE v1.1, it narrowly beats models like Anthropic's Opus 5 and OpenAI's GPT-5.6 Sol at 74.2 percent, while on ProgramBench it trails badly.
On scientifically demanding agent tasks that require expert knowledge, a clear gap to very large models remains. The technical report also acknowledges a measurable gap to leading closed systems when it comes to reading complex images.
Like many other reasoning models, the "thinking depth" can be set. Users control how thoroughly the model works through a single value, trading compute costs against accuracy. According to the report, the highest setting noticeably improves results across several benchmarks, but generates about 2.5 times as many output tokens.

Deepseek makes the model files available on Hugging Face under the open MIT license, meant as a starting point for more work on cheaper AI agents. It's also available through an API at the same prices as V4-Flash.
Deepseek had sharply improved the predecessor V4-Flash only in late July with the 0731 update. The model, with 284 billion parameters and 13 billion of them active, landed just one point behind OpenAI's GPT-5.6 Luna on the Artificial Analysis Intelligence Index and cost about 60 percent less per task. In mid-August, Deepseek took its flagship V4-Pro out of testing and raised API prices at the same time. Cache hits, meaning inputs already buffered, got six times more expensive. Chinese hacker groups have more than doubled their attacks, according to Taiwanese security firm TeamT5, since they started using Deepseek for things like exploit code and network scans.
In June, Deepseek raised about $7.4 billion in its first outside funding round at a valuation above $50 billion, and according to Reuters has now hired Chinese investment bank CITIC Securities for an IPO in China.