Next-generation AI programming model, redefining code generation boundaries
New paper by Liang Wenfeng proposes innovative Engram module to solve Transformer memory challenges, enabling model capacity without parameter bloating. This breakthrough technology expected to debut in V4 model.
DeepSeek may release new V4 model around Chinese New Year, focusing on enhanced coding capabilities. Internal tests show coding performance surpassing Claude and GPT series. V4 may adopt new pre-training framework with reasoning tech and sparse attention mechanisms.
DeepSeek to release V4 flagship model mid-February, marking architectural shift from reasoning to programming following R1. Sources indicate V4 may surpass Claude and GPT in programming tasks.
V3's core technical advantage lies in innovative MoE (Mixture of Experts) architecture. V4 will further optimize this, adopting fine-grained experts + generalist strategy to better approximate continuous multi-dimensional knowledge space.
According to The Information, DeepSeek plans to release V4 around Chinese New Year 2025. Internal tests show programming capabilities surpassing Claude and ChatGPT. New model will focus on code generation and long context processing.
DeepSeek team introduces conditional memory as supplementary sparsity dimension, implemented through Engram conditional memory module, optimizing trade-off between neural computation and static memory.
Internal tests show programming capabilities surpassing Claude and GPT series, supporting multiple programming languages and complex project development
Deep understanding of code logic and architecture design, providing high-quality code suggestions and refactoring solutions
Supports ultra-long code context, easily handling large projects and complex codebases
MoE architecture ensures efficient inference speed, significantly reducing API call costs
Supports integration with IDEs and CI/CD tools, seamlessly fitting into development workflows
Continuing DeepSeek's open source strategy, providing API and local deployment options
| Features | DeepSeek V4 | Claude | GPT-4 |
|---|---|---|---|
| Programming Capabilities | Surpassing Industry Leaders | Excellent | Excellent |
| Architecture Innovation | MoE + Engram | Transformer | Transformer |
| Cost Effectiveness | Excellent (Only 5% of GPT-4 cost) | Medium | High |
| Open Source Strategy | Fully Open Source | Closed Source | Closed Source |
| Long Context | Supports Ultra-Long Context | 200K tokens | 128K tokens |
| Inference Speed | MoE Accelerated | Medium | Medium |
July 2025 - January 2026
Liang Wenfeng's paper wins ACL2025 Best Paper Award, Engram memory module technology revealed early
January 2026
Multiple media outlets report V4 upcoming release, internal test results leaked
Mid-February 2026 (Around Chinese New Year)
DeepSeek V4 officially launches, focusing on programming capabilities
March 2026
API opens, local deployment solutions released
According to multiple sources, DeepSeek V4 is expected to be released in mid-February 2026, around Chinese New Year. This timing is similar to when DeepSeek released R1 last year.
According to internal DeepSeek employee test results, V4's performance in programming tasks indeed surpasses Claude and GPT series. Its long context processing capability and understanding of complex code structures are particularly outstanding.
V4 will continue and upgrade V3's MoE (Mixture of Experts) architecture, while introducing the new Engram conditional memory module. This "computation + memory" separation architecture solves Transformer's memory challenges, allowing models to improve performance without solely relying on parameter scaling.
Following DeepSeek's consistent strategy, V4 will likely adopt an open source model, providing API services and local deployment options. This will enable developers and enterprises to flexibly use V4's powerful programming capabilities.
V4 is optimized for programming scenarios, suitable for code generation, code review, refactoring suggestions, bug fixes, documentation generation and other development tasks. Particularly suitable for enterprise development teams that need to handle large codebases and complex projects.
Follow DeepSeek's official website and social media for the latest release information. You can also try existing DeepSeek V3 and R1 models to understand DeepSeek's technical capabilities.