DeepSeek V4 Officially Released, Ready to Dominate Global AI with Superior Performance

Liputan6.com, Jakarta - Chinese AI startup DeepSeek has officially launched its latest flagship artificial intelligence (AI) model, DeepSeek V4.

The preview version and API of the model were released on Friday, April 24, 2026, marking a significant step in the global AI technology competition.

DeepSeek V4 is specifically designed to challenge leading AI models from Western companies such as OpenAI and Anthropic, with a focus on improved reasoning capabilities and superior performance.

The launch of DeepSeek V4 was previously predicted for mid-February 2026, but finally arrived in late April.

The model comes in two main variants, DeepSeek-V4-Pro and DeepSeek-V4-Flash.

Both variants feature a 1 million-token context window, a feature that allows DeepSeek V4 to process and understand massive amounts of data simultaneously, opening up new opportunities in AI applications.

🚀 DeepSeek-V4 Preview is officially live & open-sourced! Welcome to the era of cost-effective 1M context length.🔹 DeepSeek-V4-Pro: 1.6T total / 49B active params. Performance rivaling the world's top closed-source models.🔹 DeepSeek-V4-Flash: 284B total / 13B active params.… pic.twitter.com/n1AgwMIymu

— DeepSeek (@deepseek_ai) April 24, 2026

DeepSeek V4 comes in two main versions to meet diverse user needs.

DeepSeek-V4-Pro is the larger version with 1.6 trillion parameters, designed for top-tier performance in coding benchmarking and complex reasoning tasks.

Meanwhile, DeepSeek-V4-Flash is the smaller and leaner version with 284 billion parameters.

V4-Flash is designed for low latency and cost-efficiency in real-time applications, making it a more economical and efficient choice.

Both DeepSeek V4 models feature a 1 million-token context window, allows the model to process large amounts of data, equivalent to an entire medium-sized codebase or approximately 15-20 complete novels.

DeepSeek claims that this 1 million-tokens context window is achieved with 'world-leading' cost efficiency.

DeepSeek V4 is specifically optimized for software coding and development with Repository-Level Reasoning, which enables the model to understand the entire repository structure.

This facilitates cross-file reasoning and complex bug solving, including understanding import/export relationships, tracking type definitions across modules, maintaining API signature consistency, and identifying dead code or unused dependencies.

Furthermore, V4 is capable of Multi-File Bug Diagnosis and Repair, analyzing stack traces, tracing execution paths, and proposing fixes that take the full system context into account.

The model is designed to maintain 100% logical consistency across long contexts, effectively avoiding the 'logical hallucinations' common in other models.

DeepSeek V4 introduces several fundamental architectural innovations that set it apart from previous models.

One is Manifold-Constrained Hyper-Connections (mHC), an architecture that revolutionizes the way information flows through transformer networks to helps stabilize deep signal propagation, allowing the model to maintain context across very long codebases or documents.

Another innovation is Engram Conditional Memory, a memory system that enables efficient information retrieval from very large context windows without explosive increases in computational costs.

To efficiently process 1 million tokens, V4 also replaces standard dense attention with DeepSeek Sparse Attention (DSA) and Lightning Indexer.

DeepSeek V4 also utilizes a Mixture-of-Experts (MoE) architecture with approximately 1 trillion total parameters.

However, only about 32 to 49 billion parameters are active per token, effectively keeping inference costs low.

Additionally, DeepSeek V4 utilizes the Muon Optimizer for faster and more stable convergence during the training process.

DeepSeek V4-Pro is claimed to achieve 'top-tier' performance in coding benchmarks and significantly bridges the gap with leading closed-source models in reasoning and agent tasks.

In the MMLU-Pro benchmark, DeepSeek V4-Pro is said to be on par with OpenAI's GPT-5.4, although slightly behind Google's Gemini-3.1-Pro and Anthropic's Claude Opus 4.6.

In SWE-bench testing, V4 aims to solve over 80.9% of complex real-world problems.

The model also significantly outperforms other open-source models in world knowledge benchmarks and is only slightly outperformed by the top closed-source model, Google's Gemini-Pro-3.1.

The MoE V4 architecture offers approximately 40% faster inference speeds than Claude 3.5 Sonnet at a lower cost.

DeepSeek V4 is designed to be highly cost-effective.

The API price is estimated at approximately $0.14 per 1 million input tokens, which is approximately 20-50 times cheaper than the Western frontier model.

The cost per task is estimated at only $0.03, compared to $0.72 for Claude.