{"id":2616,"date":"2026-04-24T04:47:07","date_gmt":"2026-04-24T04:47:07","guid":{"rendered":"https:\/\/deepinsightai.io\/?p=2616"},"modified":"2026-04-24T04:47:09","modified_gmt":"2026-04-24T04:47:09","slug":"deepseek-v4","status":"publish","type":"post","link":"https:\/\/deepinsightai.io\/ja\/deepseek-v4\/","title":{"rendered":"DeepSeek V4 Is Here: 1M Context Becomes Standard, Challenging Top Closed-Source AI"},"content":{"rendered":"

DeepSeek V4, which made the whole world wait bitterly until April, has finally arrived!<\/p>\n\n\n\n

Just now, DeepSeek V4 really came!<\/p>\n\n\n\n

Today, DeepSeek, the one that once broke the dominance of closed-source models almost by itself and proved that DeepSeek starts updating frequently<\/a> to shift industry dynamics, has officially announced to global developers with the preview version of the DeepSeek-V4 series\u2014<\/p>\n\n\n\n

The civilian era of million-level context (1M Context), and a new peak in open-source Agent capabilities, world knowledge, and reasoning performance, has arrived.<\/p>\n\n\n\n

DeepSeek V4 has once again achieved leadership in China and in the open-source field, mirroring the rapid rise we’ve seen as Qwen 3.6<\/a> pushed boundaries.<\/p>\n\n\n\n

The technical report of V4 has also been released at the same time.<\/p>\n\n\n\n

\"deepseek<\/figure>\n\n\n\n


Paper address: https:\/\/huggingface.co\/deepseek-ai\/DeepSeek-V4-Pro\/blob\/main\/DeepSeek_V4.pdf<\/a><\/p>\n\n\n\n

DeepSeek V4 Models Overview: V4-Pro vs V4-Flash<\/h2>\n\n\n\n

DeepSeek V4-Pro: Flagship Model Specifications<\/h3>\n\n\n\n

The DeepSeek-V4 series includes two versions: the performance monster DeepSeek-V4-Pro, with 1.6T total parameters and 49B activated parameters.<\/p>\n\n\n\n

DeepSeek V4-Flash: Efficient and Cost-Optimized Version<\/h3>\n\n\n\n

DeepSeek-V4-Flash is designed specially for high efficiency and economy, with 284B total parameters and 13B activated parameters.<\/p>\n\n\n\n

\"deepseek<\/figure>\n\n\n\n

It can be said that DeepSeek-V4-Pro has reached a new peak for open-source models, benchmarking against the world\u2019s top closed-source level.<\/p>\n\n\n\n

DeepSeek V4-Pro Performance: Agent, Knowledge, and Reasoning<\/h2>\n\n\n\n

DeepSeek V4 Agent Capability Breakthrough<\/h3>\n\n\n\n

First, V4-Pro has achieved a leap-forward breakthrough in Agent capabilities, and its Agentic Coding<\/a> level firmly ranks first in the open-source world.<\/p>\n\n\n\n

Actual test feedback shows that its coding experience has already surpassed Sonnet 4.5, and its delivery quality is catching up with Opus 4.6<\/a> (non-thinking mode). It has now become the preferred model for internal Agent programming<\/a> in the company.<\/p>\n\n\n\n

DeepSeek V4 World Knowledge Benchmark Performance<\/h3>\n\n\n\n

Second, it has deep world knowledge reserves.<\/p>\n\n\n\n

In the knowledge evaluation dimension, V4-Pro is significantly ahead of similar open-source products, and the gap with the closed-source benchmark Gemini-Pro-3.1 has been reduced to a very small range.<\/p>\n\n\n\n

DeepSeek V4 Reasoning and STEM Capabilities<\/h3>\n\n\n\n
\"deepseek<\/figure>\n\n\n\n

Also, it has top-tier logical reasoning performance.<\/p>\n\n\n\n

In hard-core areas such as mathematics, STEM<\/a>, and high-difficulty competitive programming, V4-Pro not only tops the open-source community, but also already has real combat competitiveness to challenge the strongest closed-source models in the world.<\/p>\n\n\n\n

DeepSeek V4 Architecture Innovations: CSA, HCA, and mHC<\/h2>\n\n\n\n

Supporting these two models in looking down on the field are the \u201cthree great skills\u201d of the underlying technology:<\/p>\n\n\n\n

Hybrid Attention in DeepSeek V4 (CSA + HCA)<\/h3>\n\n\n\n

DeepSeek-V4 did not blindly increase hardware investment, but creatively designed a hybrid attention architecture.<\/p>\n\n\n\n

Compressed Sparse Attention (CSA) compresses the KV cache along the token dimension and combines it with DSA sparse attention; Heavy Compressed Attention (HCA) carries out even more extreme compression to maintain dense computation.<\/p>\n\n\n\n

This \u201clong and short combination\u201d strategy greatly reduces computation and memory requirements when the model handles million-word context.<\/p>\n\n\n\n

Manifold-Constrained Hyper-Connection (mHC) in DeepSeek V4<\/h3>\n\n\n\n

To improve the stability of signal propagation and strengthen model expressiveness, V4 introduces the mHC structure, upgrading the traditional residual connection.<\/p>\n\n\n\n

Muon Optimizer in DeepSeek V4 Training<\/h3>\n\n\n\n

The new Muon optimizer is introduced, making the training process not only converge faster, but also more stable.<\/p>\n\n\n\n

It is exactly these structural innovations that allow DeepSeek-V4 to achieve a qualitative leap in inference efficiency.<\/p>\n\n\n\n

In the extreme scenario of 1 million token context, DeepSeek-V4-Pro\u2019s single-token inference computation is only 27% of the previous generation, and KV cache usage is reduced to an astonishing 10%.<\/p>\n\n\n\n

DeepSeek V4-Flash Performance: Speed, Cost, and Use Cases<\/h2>\n\n\n\n

Compared with the Pro version, the Flash version is a faster and more efficient economical choice.<\/p>\n\n\n\n

DeepSeek V4-Flash Reasoning vs Pro Comparison<\/h3>\n\n\n\n

Although it is slightly weaker than the Pro version in the depth of world knowledge, DeepSeek-V4-Flash keeps a logical reasoning level close to it.<\/p>\n\n\n\n

DeepSeek V4 API Cost and Efficiency Advantages<\/h3>\n\n\n\n

Benefiting from a more streamlined parameter scale and activation mechanism, it can provide users with API access that responds faster and costs less.<\/p>\n\n\n\n

DeepSeek V4 Agent Task Performance Differences<\/h3>\n\n\n\n

When handling basic Agent tasks, V4-Flash performs almost the same as the Pro version, but when facing extremely complex tasks, there is still room for further progress.<\/p>\n\n\n\n

DeepSeek V4 Long Context Breakthrough: 1M Tokens Becomes Standard<\/h2>\n\n\n\n
\"DeepSeek<\/figure>\n\n\n\n
\"DeepSeek<\/figure>\n\n\n\n

DeepSeek-V4 introduces a revolutionary attention mechanism.<\/p>\n\n\n\n

Through efficient compression in the Token dimension and combining DSA sparse attention (DeepSeek Sparse Attention) technology, it achieves world-leading long-text processing capability.<\/p>\n\n\n\n

DeepSeek V4 1M Context as Default Configuration<\/h3>\n\n\n\n

Starting today, 1M (1 million tokens) ultra-long context will become the standard configuration of DeepSeek official services.<\/p>\n\n\n\n

DeepSeek V4 vs Previous Models in Memory and Compute<\/h3>\n\n\n\n

This innovation greatly cuts dependence on computing resources and memory.
Changes in computation and memory capacity of DeepSeek-V4 and DeepSeek-V3.2 with context length<\/p>\n\n\n\n

DeepSeek V4 Agent Optimization and Ecosystem Integration<\/h2>\n\n\n\n

DeepSeek V4 Integration with Agent Frameworks<\/h3>\n\n\n\n

DeepSeek-V4 has carried out deep adaptation for mainstream Agent ecosystems such as Claude Code, OpenClaw<\/a>, OpenCode, and CodeBuddy.<\/p>\n\n\n\n

DeepSeek V4 Use Cases: Code and Document Generation<\/h3>\n\n\n\n

In scenarios such as code writing and automated document generation, its output efficiency has been significantly improved.
Example of PPT pages automatically generated by V4-Pro under a specific Agent framework<\/p>\n\n\n\n

DeepSeek V4 API Update and Model Migration Guide<\/h2>\n\n\n\n

How to Use DeepSeek V4 API (model_name Setup)<\/h3>\n\n\n\n

For developers, the good news is: the API has already gone online at the same time!<\/p>\n\n\n\n

You only need to simply modify model_name to access these two new flagships:<\/p>\n\n\n\n