In Brief: DeepSeek has released preview weights for V4-Pro (1.6T parameters, 49B active) and V4-Flash (284B / 13B active) under the MIT licence — topping all open-source models on coding and maths benchmarks, and self-reported within "3 to 6 months" of GPT-5.4 and Gemini 3.1-Pro.
China's DeepSeek posted preview weights for DeepSeek-V4-Pro and DeepSeek-V4-Flash to Hugging Face on Friday, April 24, roughly one year after R1 upended the AI industry's assumptions about the cost of building frontier-level AI. Both models carry a permissive MIT licence and a one-million-token context window, making them immediately deployable for enterprises and researchers at a fraction of what closed-source alternatives charge.
V4-Pro, at 1.6 trillion total parameters with 49 billion active at inference, claims the top position among open-weight models on coding and mathematics benchmarks. V4-Flash — 284B total, 13B active — is designed for speed and cost-efficient inference. On the Artificial Analysis Intelligence Index, V4-Pro scores 52, ranking it the #2 open-weight reasoning model globally behind Kimi K2.6, while V4-Flash scores 47. DeepSeek's own technical documentation concedes that V4-Pro falls "marginally short" of GPT-5.4 and Gemini 3.1-Pro, estimating the remaining gap at three to six months of development. API pricing for Flash starts at $0.14 per million input tokens; Pro runs $1.74 per million — a fraction of comparable closed-source rates.
The timing is strategically pointed. V4 arrives just days after OpenAI released GPT-5.4 "Thinking" and Google confirmed Gemini 3.1's rollout, at a moment when the Western frontier is absorbing record levels of capital investment. DeepSeek's repeated ability to produce near-frontier results at low training and inference cost continues to challenge the assumption that dominance in AI requires Silicon Valley-scale funding — a point that landed with particular force when R1 launched in early 2025 and that V4 reinforces.
The preview label is worth noting: full weights are not yet released, and independent benchmark reproduction is still underway. The key things to watch are whether third-party evaluators confirm the coding and reasoning claims at scale, and whether DeepSeek follows up with a distilled or quantised version optimised for consumer hardware — its typical playbook after a flagship launch.