DeepSeek previews a new type of AI that ‘closes the gap’ with borderline species

Chinese AI lab DeepSeek has launched two preview versions of its newest language, DeepSeek V4the much-anticipated update to last year’s V3.2 version and its sequels R1 consider the model that he took AI world is a storm.

The company says that both DeepSeek V4 Flash and V4 Pro are hybrid professional models with windows of 1 million tokens each – enough to allow large codebases or documents to be used quickly. A mixed method with professionals involves starting multiple stages for each job to reduce the estimated costs.

The Pro version has 1.6 trillion parts (49 billion active), making it the largest visible version available, surpassing Kimi K 2.6’s Moonshot AI (1.1 trillion), MiniMax’s M1 (456 billion), and surpassing DeepSeek V3.2 (671 billion). Small, V4 Flash has 284 billion parts (13 billion active).

DeepSeek claims that both models are more effective and efficient than DeepSeek V3.2 due to the revision of the architecture, and they have almost “closed the gap” with the current leading models, open and closed, in terms of considerations.

The company says its new version of V4-Pro-Max outperforms its open-source counterparts on benchmarks, and outperforms OpenAI’s GPT-5.2 and Gemini 3.0 Pro on other tasks. As for the curves, DeepSeek said that all of V4’s performance is “the same as GPT-5.4.”

However, the models seem to fall a little behind in cognitive tests, especially OpenAI’s GPT-5.4 and Google’s latest Gemini 3.1 Pro. The lag shows a “developmental pattern that follows the most advanced samples for about 3 to 6 months,” the lab wrote.

Both V4 Flash and V4 Pro support themselves, unlike many of their closed counterparts, which provide support for understanding and creating audio, video, and images.

In particular, DeepSeek V4 is cheaper than any other border type available today. The limited version of V4 Flash costs $0.14 per million input tokens and $0.28 per million output tokens, downloading GPT-5.4 Nano, Gemini 3.1 Flash, GPT-5.4 Mini, and Claude Haiku 4.5. The larger V4 Pro version, meanwhile, costs $0.145 per million input tokens and $3.48 per million output tokens, dwarfing Gemini 3.1 Pro, GPT-5.5, Claude Opus 4.7, and GPT-5.4.

The launch comes just a day after the US launch the accused China stole American AI labs’ IP on companies using thousands of proxy accounts. DeepSeek itself has been accused by Anthropic and OpenAI of “distilling,” literally copying, their AI models.