DeepSeek-V4 Preview: Powerful Open-Weights Model with API Access

The Chinese AI company DeepSeek made headlines last year by topping the App Store charts and outperforming competitors like ChatGPT. With the launch of its latest model, DeepSeek-V4, the company aims to set new benchmarks in AI capabilities.

Users with a DeepSeek account can test the new model at chat.deepseek.com/. Two versions are available: Expert and Instant. The Expert model boasts an impressive 1.6 trillion parameters, while the Instant model features a still considerable 284 billion parameters.

Notably, DeepSeek-V4 is an open-weights model, allowing users to download it from Hugging Face and operate it on their own hardware, though significant computational resources are required to effectively run these models.

The Expert (Pro) model utilizes 49 billion active parameters, whereas the Instant (Flash) model uses 13 billion active parameters—these are the parameters that need to fit into VRAM. Although transferring parameters between VRAM and system memory can slow down token generation, open-weight models enable the community to develop quantized and distilled versions that can run on consumer-grade hardware.

The DeepSeek-V4-Flash model, which powers the Instant mode, reportedly has reasoning capabilities that closely match those of the V4-Pro model and performs efficiently in basic agentic tasks. Both models support a context window of 1 million tokens.

According to internal testing, DeepSeek-V4-Pro excels in reasoning for mathematical, STEM, and coding tasks among open-weight models and competes closely with proprietary models. The V4-Pro model also has superior world knowledge, trailing only behind Google’s Gemini-3.1-Pro.

DeepSeek-V4 integrates seamlessly with popular tools such as Claude Code, OpenClaw, and OpenCode. For those unable to run the model locally, DeepSeek offers an accessible API, with pricing details provided below.