By RXA AI News Desk
November 11, 2026
Continuing its aggressive disruption of the open-weight market, China-based lab DeepSeek has officially released DeepSeek V4.1 Flash under the commercially friendly MIT license. Trained on over 45 trillion tokens, the new Mixture-of-Experts (MoE) model features 16 billion active parameters per routing step out of a staggering 1.6 trillion total parameter pool. In an unexpected twist, independent evaluations show V4.1 Flash outperforming its own predecessors on dense multi-step analytical reasoning while supporting a massive 1-million-token context window.
The real breakthrough, however, lies in its hardware optimization. DeepSeek engineers confirmed that V4.1 Flash was compiled and fully optimized using completely non-NVIDIA accelerator hardware. By leveraging alternative domestic GPU clusters, DeepSeek has demonstrated that state-of-the-art inference can survive a decoupled supply chain. The model introduces a proprietary "engram memory layer" that slashes token generation costs down to a fraction of western proprietary equivalents, setting off a massive localized deployment race among global startups.