The lineup includes DeepSeek-V4-Pro with 1.6 trillion total parameters and V4-Flash at 284 billion, both released under the MIT license on Hugging Face and supporting base and instruct variants. The models integrate key innovations such as token-wise compression and Sparse Attention, enabling 1M context windows across web chat, app, and API environments while maintaining cost efficiency. Pricing positions the models far below rivals, with V4-Flash available at $0.028 per million input tokens on cache hits.
Benchmark performance underscores the ambition of the release. V4-Pro-Max demonstrates competitive results against top closed models, achieving 87.5% on MMLU-Pro, 93.5% on LiveCodeBench, and strong performance across coding tasks. Despite these gains, some in the community note that it remains slightly behind leading systems such as Claude Opus in real-world application performance, even as it is widely hailed as open-source royalty.
The release is accompanied by immediate availability for users via web chat, app, and API access, alongside an administrative update indicating that older models will be deprecated on July 24, 2026.
The unveiling of the V4 Series reinforces DeepSeek’s strategic push to redefine open-source AI capabilities, combining scale, affordability, and advanced architecture, while intensifying competition with established closed-model leaders.
About The Author
Comment List