DeepSeek-V4 Preview is now live and open-sourced, the Chinese firm has said. The company is pitching this model as cost-effective 1M context length capable. The series includes two variants, Pro and Flash, and both are available on the DeepSeek website via Instant Mode and Expert Mode. It also said that the API has been updated and is available today.
Both Flash and Pro support 1M context lengths but the real difference is in the total parameters and active parameters. Pro has 1.6 trillion total parameters and 49 billion active parameters, while Flash has a smaller 284 billion parameters and 13 billion parameters.
According to DeepSeek, the Pro variant delivers enhanced agentic capabilities, has rich world knowledge that surpasses all current open models and is behind on Gemini 3.1 Pro, and it displays world-class reasoning that beats all current open models in Math, STEM, and coding, and rivals the top closed-source models.
With the Flash model, you get reasoning capabilities that closely approach the Pro variant, it performs on par with Pro for simple Agent tasks, and it’s more cost-effective with its API pricing.
In terms of API pricing, Flash inputs cost as low as $0.028 or $0.14 if there is a cache miss, and an output price of $0.28. For the Pro model, the respective prices are $0.145 and $1.74 for input (cache hit vs cache miss), and an output cost of $3.48.
To begin chatting with these models now, head over to https://chat.deepseek.com/. If you want to check out the Open Weights for these models head here, and if you’d like to take a deeper dive into the technical details then check out the Tech Report.
Source: DeepSeek
5 Comments
Load the comments and join the conversation!
Read the comments, ask the editors questions, show respect and join the conversation.