Chinese AI company DeepSeek has released a preview of its latest model, called DeepSeek-V4, more than a year after it gained attention by climbing to the top of App Store charts and competing with services such as ChatGPT.
The new release includes two versions, Expert and Instant, available through DeepSeek’s platform for users with an account.
DeepSeek said the V4-Flash model, which powers Instant mode, offers reasoning performance close to V4-Pro while matching it on simpler agent-based tasks. Both models support a 1 million token context window.
DeepSeek said V4-Pro, used in Expert mode, delivers strong reasoning performance for mathematics, STEM, and coding tasks among open-weight models.
The company also claimed it rivals closed-source competitors. In world knowledge benchmarks, DeepSeek said V4-Pro trails only Google’s Gemini-3.1-Pro.
DeepSeek-V4 Expert uses a 1.6 trillion parameter architecture. The Instant version uses a smaller 284 billion parameter model.
According to the company, the Expert model has 49 billion active parameters, while Instant uses 13 billion active parameters. Active parameters are the parts of the model that need to fit into VRAM during use. DeepSeek noted that moving parameters between VRAM and system memory can slow token generation.
DeepSeek-V4 is being offered as an open-weight model. That means users can download it from Hugging Face and run it on their own hardware, though the company noted that substantial hardware resources are needed for full performance.
The company added that open-weight availability allows the community to create quantized and distilled versions that may run on consumer-level hardware.
DeepSeek-V4 can be used with tools including Claude Code, OpenClaw, and OpenCode. Users who cannot run the models locally can access them through DeepSeek’s API. The company has also published pricing for V4 Pro and Flash API access.