DeepSeek Unveils New Model and Search API for AI Agents

DeepSeek has unveiled an updated version of its V4.1-Flash-1017 language model and launched a proprietary Search API for AI agent integration.

The new model supports a context window of up to 1 million tokens, features open weights under the MIT license, and is accessible via API. The version focuses on enhancing coding capabilities, logical reasoning, and autonomous agent performance.

Benchmark results demonstrate the model's competitiveness against market leaders.

DeepSeek achieves high scores in specialized benchmarks:

1. DeepSWE: 77.0 points (outperforming Claude Opus 5 and GPT-5.6 Sol)
2. Automation-Bench: 60.5 points (better than Opus 5)

While Anthropic leads in Terminal-Bench 4.0 and ProgramBench tasks, DeepSeek already surpasses competitors in agentic and coding challenges.

Simultaneously, the Search API was launched, capable of processing hundreds of millions of queries daily.

The API offers two tools for external developers and AI agents:

1. Search — internet search with AI-optimized results
2. Fetch — extraction of content from found pages

Pricing is set at $0.75 per 1,000 calls for each tool. Search is free for specific server-side integrations until October 31.

The technology is compatible with external agents, including Claude Code and Codex, enabling DeepSeek to compete with Google, Bing, and other search engines.

Enthusiasts have already deployed a quantized version of the model on four NVIDIA DGX Spark units via vLLM, with a checkpoint size of 510.3 GB.

Documentation and access to the models are available at the following links:

• DeepSeek News
• Model on Hugging Face
• API Updates