DeepSeek-V4 Preview arrives with open weights, 1 million-token context and a new API pricing play
DeepSeek has launched a preview of DeepSeek-V4, its latest large language model, offering both downloadable open weights and paid API access.
The release positions the Chinese company to compete more directly with leading proprietary and open models as demand grows for controllable, self-hosted systems.
The preview is available through DeepSeek’s chat interface to account holders, with two modes aimed at different needs. DeepSeek labels them Expert and Instant, corresponding to V4-Pro and V4-Flash in its documentation.
Two models, different trade-offs
DeepSeek says V4-Pro is a 1.6 trillion-parameter model, while V4-Flash is a 284 billion-parameter alternative designed for faster, lighter use. Both are built with a mixture-of-experts approach, meaning only a subset of parameters are active for each token.
According to the company, V4-Pro uses 49 billion active parameters and V4-Flash uses 13 billion active parameters, which is the portion that must fit in VRAM for efficient generation. DeepSeek notes performance can drop if systems rely heavily on swapping between VRAM and system memory.
Why open weights still matter
By releasing open weights, DeepSeek enables researchers and developers to download the models and run them on their own hardware, subject to practical compute limits. The approach also allows the community to create quantized or distilled variants that may run on more accessible machines.
DeepSeek says both models support a 1 000 000-token context window, a specification that can benefit long-document workflows and multi-step agent tasks. In practice, real-world throughput and cost will depend on deployment choices and hardware constraints.
Benchmarks, tools, and API access
DeepSeek claims V4-Flash delivers reasoning close to V4-Pro and performs similarly on simpler agent-style tasks, based on its internal testing. It also says V4-Pro leads among open-weight models in math, STEM, and coding reasoning, and is competitive with several closed systems.
The company is also pushing integration, saying DeepSeek-V4 can be used with popular developer tools such as Claude Code, OpenClaw, and OpenCode. For teams that cannot self-host, DeepSeek is offering API access with published pricing for both V4-Pro and V4-Flash.
