China’s DeepSeek has launched preview versions of its flagship V4 model, calling it the most powerful open-source AI platform in the world. The release marks one year since the original DeepSeek breakthrough that rattled global markets.
Two Variants
- DeepSeek V4-Pro — Mixture-of-Experts model with 1.6 trillion total parameters, 49 billion activated
- DeepSeek V4-Flash — 284 billion total parameters, 13 billion activated
- Both models support a 1 million-token context window — large enough to ingest entire codebases or long documents in a single prompt
Hybrid Attention Architecture
DeepSeek highlights a new technique it calls Hybrid Attention Architecture, designed to improve recall across long conversations and agentic workflows where models historically lose track of context.
Open-Source on Hugging Face
Like previous DeepSeek releases, V4 is fully open-source on Hugging Face — developers can download the weights, run them locally and modify them. Technical reports are also published. The release is accompanied by what the company describes as rock-bottom inference pricing.
Huawei Chip Bombshell
The launch is closely integrated with Huawei’s chip stack — a deliberate signal of Chinese AI self-sufficiency at a time when Washington is tightening export controls on advanced GPUs. DeepSeek positions V4 as a direct challenge to OpenAI’s GPT-5.4 and Anthropic’s Claude Opus 4.6, with strong reported performance in coding benchmarks, knowledge processing and agentic tasks.
Follow Vibes Uncut Media for AI infrastructure coverage.
Leave a Reply