Overview
- DeepSeek, which opened a public preview Friday, posted two Mixture‑of‑Experts language models on Hugging Face and made them available in its chat and API.
- V4 Pro lists 1.6 trillion total parameters with 49 billion active at inference, and V4 Flash lists 284 billion total with 13 billion active, both targeting long tasks with a one‑million‑token context window.
- The company says its MoE design cuts compute and memory needs by activating only a small set of experts per request, which aims to speed replies and lower cost without dropping quality.
- DeepSeek reports V4 Pro tops current open models and comes close to leading closed systems, with its own tests placing it near Google’s Gemini 3.1 Pro, though analysts say independent checks are still needed.
- Huawei said its Ascend chips run the new models, a shift that reduces reliance on Nvidia as the U.S.–China tech rivalry sharpens after prior DeepSeek adoption raised privacy and censorship concerns and U.S. firms alleged model distillation.