EDMONTON, CANADA - JANUARY 28: A woman holds a cell phone in front of a computer screen displaying the DeepSeek logo, on January 28, 2025, in Edmonton, Canada. (Photo by Artur Widak/NurPhoto via Getty Images)
NurPhoto via Getty Images
DeepSeek V4, the long awaited update from DeepSeek, arrives at a fiercely competitive moment in the AI models race, when Open AI’s GPT 5.5 and Anthropic’s Opus 4.7 have just launched one after the other.
As a unique believer in open sourced tools, DeepSeek has impressed developers with its cost-efficiency rather than its raw scale. The preview release includes two mixture-of-experts models with 1 million–token context window: DeepSeek-V4-Pro, with 1.6 trillion total parameters and 49 billion activated parameters, and DeepSeek-V4-Flash, with 284 billion total parameters and 13 billion activated parameters.
Long-context agents, coding assistants, research tools and enterprise copilots all face the same bottleneck: every newly generated token may need to refer back to a growing history of documents, code, tool calls and intermediate reasoning. DeepSeek’s technical report demonstrates that its V4 models addresses this problem through architectural compression rather than simply asking users to pay for more compute.
DeepSeek V4’s most important architectural change is a hybrid attention design that combines “compressed sparse attention” with “heavily compressed attention.” It means that the model does not store and scan every previous token in the same expensive way. CSA compresses groups of key-value entries and then selects the most relevant compressed blocks. HCA compresses even more aggressively, allowing dense attention over a much shorter memory stream.
This matters because attention is one of the main cost drivers in long-context AI. As context length grows, conventional attention becomes increasingly expensive in both computation and memory. DeepSeek’s hybrid attention design treats long context as an engineering problem of memory hierarchy. Some information needs fine-grained local attention. Some can be compressed. By combining these modes, V4 turns million-token context into a more practical capability.
Earlier this year, DeepSeek researchers published a paper proposing Engram, a conditional memory module that advances reasoning efficiency by structurally separating static knowledge retrieval from dynamic computation.
Lower inference cost changes who can experiment. When long-context reasoning becomes cheaper, more developers can build agents that read full repositories, analyze long legal records, compare multi-document financial filings or operate across extended tool-use sessions. This expands the design space beyond chatbot prompts.
For startups, DeepSeek V4 lowers the cost of trying ambitious applications. For enterprises, it makes large-context workflows more realistic. For open-source developers, it provides a technical recipe: combine mixture-of-experts sparsity, long-context compression, low-precision inference, custom kernels and post-training for agentic tasks.
DeepSeek V4 is also notable because the technical report makes explicit suggestions on hardware design. The team argues that future hardware should optimize for the ratio between computation and communication rather than blindly increasing bandwidth.
Reuters also reported that DeepSeek V4 has been adapted to run on Huawei’s Ascend chips, and that Huawei said its Ascend 950-based supernode clusters fully support the V4 series. This makes V4 part of a larger hardware story. The AI race is moving from model weights to full-stack co-design, where models, kernels, memory systems, interconnects and chips co-evolve.
The most important consequence of DeepSeek V4 may be economic. When the cost of long-context reasoning falls, AI use cases that once looked too expensive become more plausible. Full-codebase agents, long-horizon research assistants, document-heavy legal workflows, financial diligence tools, scientific literature review systems and enterprise knowledge agents all benefit from cheaper memory and cheaper inference.
This means that DeepSeek V4 reframes the AI race. If DeepSeek can deliver strong open models with lower memory and compute requirements, closed-source leaders will face more pressure to justify premium pricing. Open-source competitors will face pressure to match V4’s efficiency techniques.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。