DeepSeek V4: China's Response to GPT-5.5 and Claude

DeepSeek V4: A New Era for Chinese AI

A year and a half after shaking up the artificial intelligence sector, DeepSeek, the Hangzhou-based startup, is back with a new version of its model, DeepSeek V4. At the beginning of 2025, the company had already surprised the world with high-performing models at development costs significantly lower than their American counterparts. Since then, DeepSeek had kept a low profile, but on April 24, it announced the release of DeepSeek V4 in open-source pre-release, promising to compete with the very recent GPT-5.5.

Two Distinct Versions for Varied Needs

DeepSeek V4 comes in two versions, each with its own architecture:

• DeepSeek V4-Pro: This impressive model boasts 1.6 trillion parameters, of which 49 billion are active. It is designed for advanced applications, with agentic capabilities superior to those of the previous version. DeepSeek is already using it for its internal agentic coding processes and has integrated it with tools such as Claude Code and OpenCode.

• DeepSeek V4-Flash: With 284 billion parameters and 13 billion active, this version is optimized for fast and economical use. Although less powerful than the V4-Pro, it offers comparable performance for simple agentic tasks.

Both models support a context of one million tokens, a capacity among the most competitive on the market. DeepSeek achieved this performance through an innovative attention architecture, including token compression and the DSA (DeepSeek Sparse Attention) mechanism, thereby reducing computational and memory costs.

Concrete Advances of DeepSeek V4

DeepSeek highlights three key areas where the V4-Pro excels compared to its predecessor V3:

• Agentic Capabilities: On the Codeforces platform, V4-Pro outperforms GPT-5.4 and Gemini-3.1-Pro. Regarding SWE Verified, which assesses the autonomous resolution of software tickets, all three models achieve nearly 80% success.

• Mathematical and Scientific Reasoning: DeepSeek claims that V4-Pro surpasses current open-source models in mathematics, STEM, and coding, competing with the best proprietary models. On the Apex Shortlist, it scores 90.2, exceeding Claude Opus 4.6 and GPT-5.4.

• General Knowledge: V4-Pro outperforms Claude Opus 4.6 and GPT-5.4 on SimpleQA Verified, although it still trails behind Gemini-3.1-Pro in this regard.

A detailed technical report is available on the model's Hugging Face page for those who wish to delve deeper.

Immediate and Free Availability of DeepSeek V4

DeepSeek V4 is available for free starting today through several channels:

• On the chat.deepseek.com interface, users can choose between Expert Mode (V4-Pro) and Instant Mode (V4-Flash).

• The DeepSeek API is also available, with model identifiers deepseek-v4-pro and deepseek-v4-flash. It is compatible with OpenAI ChatCompletions and Anthropic APIs.

It is worth noting that the older models deepseek-chat and deepseek-reasoner will be permanently retired on July 24, 2026.