DeepSeek released V4, an open-weights LLM with 284 billion parameter Flash MoE and 1.6 trillion parameter variants, trained on 33 trillion tokens. The models claim performance rivaling proprietary frontier models while reducing inference costs through hybrid attention mechanisms and mixed FP8/FP4 precision. V4 now supports Huawei Ascend NPUs and is available via Hugging Face, API, and web service.
Models
DeepSeek's new models are so efficient they'll run on a toaster ... by which we mean Huawei's NPUs
DeepSeek's open-weights V4 matches frontier model performance while slashing inference costs through novel efficiency techniques, now optimized for Huawei's Ascend NPUs—a major competitive threat to proprietary incumbents.
Friday, April 24, 2026 12:00 PM UTC2 MIN READSOURCE: The RegisterBY sys://pipeline
Tags
models
/// RELATED
Research3d ago
Reducing ML-KEM-768 encapsulation key sizes by 24 octets
Bit-packing optimization trims ML-KEM-768 post-quantum cryptography encapsulation keys by 24 octets, enabling better UDP packet alignment for practical PQC deployment.
Strategy4d ago
Artemis III aims for 'late 2027' for Earth orbit demonstration
NASA targets late 2027 for Artemis III Earth orbit demonstration of SpaceX and Blue Origin landers, setting up a 2028 lunar landing attempt with interoperability testing in between.