DeepSeek Previews V4 Model for Agentic AI Tasks
DeepSeek has released open-source preview versions of V4, with Pro and Flash models designed for long-context reasoning, coding and agent workflows.
Topics
News
- DeepSeek Previews V4 Model for Agentic AI Tasks
- Reliance Meta AI Venture Names Parminder Singh CEO
- Meta Layoffs, Microsoft Buyouts Signal AI Reset
- Employers Turn to Certifications as AI Skills Gaps Bite
- Nasscom Names Fractal's Velamakanni as Chairperson
- OpenAI Releases GPT 5.5 for Coding, Enterprise Work
Chinese AI startup DeepSeek on Friday released preview versions of its V4 model, more than a year after its R1 reasoning model drew global attention for delivering competitive performance at lower cost.
The Hangzhou-based company said DeepSeek-V4 is live and open-sourced, with a 1 million-token context length. It is available in two versions: DeepSeek-V4-Pro, with 1.6 trillion total parameters and 49 billion active parameters, and DeepSeek-V4-Flash, with 284 billion total parameters and 13 billion active parameters.
DeepSeek described Flash as the faster and more economical version.
The company said the models are available through its chat interface and API, with support for both OpenAI-compatible and Anthropic-compatible formats.
DeepSeek’s API documentation says V4-Pro and V4-Flash support tool calls, JSON output and a 1 million-token context window.
DeepSeek is also positioning V4 for agentic and coding workflows. Its documentation includes guidance for integrating the model with coding agents such as Claude Code and OpenCode.
Wei Sun, principal AI analyst at Counterpoint Research, said early benchmark data suggest the model could deliver strong agent capabilities at lower cost. “V4’s benchmark profile suggests it could offer excellent agent capability at significantly lower cost,” Sun said.
The release adds to competition in the global AI market, where companies are seeking to improve performance while reducing the cost of deploying models.
