DeepSeek Unveils V4 Model With One Million Token Context

Chinese artificial intelligence company DeepSeek has launched its latest large language model, DeepSeek-V4, claiming a leap forward in processing extended text contexts and improving reasoning abilities. The model can handle up to one million tokens—individual units of data in AI systems—substantially surpassing most current open-source alternatives.

DeepSeek-V4 is offered in two versions: DeepSeek-V4-Pro and DeepSeek-V4-Flash. The V4-Pro version features 1.6 trillion total parameters and 49 billion active parameters, positioning it as one of the largest open-source models. V4-Flash provides a lighter, more economical alternative with 284 billion total and 13 billion active parameters.

Benchmarks shared by DeepSeek indicate that V4-Pro matches or outperforms leading open-source large language models in world knowledge and reasoning tasks, while closely approaching the performance levels of Google’s Gemini-Pro-3.1, a prominent closed-source model. However, in handling very long text strings, it reportedly trails Anthropic’s Claude Opus 4.6.

A notable feature of DeepSeek-V4-Pro is the “maximum reasoning effort mode,” designed to further enhance its reasoning and knowledge extraction capabilities. This follows a previous market impact created by the company’s R1 release, which competed favorably with OpenAI’s ChatGPT at a lower cost point.

DeepSeek confirms that V4 models are compatible with Nvidia and Huawei chips, though it has not disclosed details about its training hardware. This compatibility is significant, given the growing restrictions from the United States on exporting advanced semiconductors to China, a factor affecting the domestic AI sector’s access to high-end computing resources.

With a maximum output of 384,000 tokens per completion, DeepSeek-V4 boasts an exceptionally large working memory compared to many of its competitors. For context, a token represents roughly four characters of text, and increasing a model’s context window is vital for tasks such as summarizing extensive documents or sustaining longer conversations.

The model’s capabilities may pave the way for more complex applications across natural language processing tasks, such as drafting lengthy documents, code analysis, and maintaining intricate conversational threads. DeepSeek intends for this advancement to mark the beginning of what it describes as “the era of million-length contexts” among next-generation language models.

The DeepSeek-V4 suite is open-sourced and available on platforms such as Hugging Face, facilitating evaluation and integration by global developers and researchers.

Reference: dataconomy.com