Chinese AI startup DeepSeek has released a preview of its flagship large language model, V4, boasting significantly extended context windows and support for Huawei NPUs. The model is open-weight, allowing developers and organizations to run it locally without relying on proprietary APIs.
"This research leaves the lab and enters your life sooner than you think," the company stated.
Key features of the V4 preview include a new architectural design that efficiently handles large amounts of text, enabling much longer context windows than previous versions. While the model shows promise, independent benchmark results have not yet been published to verify performance claims against leading proprietary models like GPT-4 and Claude.
Nevertheless, the open-source nature of DeepSeek V4 could democratize access to powerful AI for entities that cannot afford expensive proprietary APIs. This aligns with DeepSeek's history of releasing capable models at low cost.
The preview was announced on April 26, 2026, with the source cited as MIT Technology Review.