DeepSeek released V4 Pro and Flash, frontier-class open models with up to 1.6T parameters and 1M-token context using novel Compressed Sparse and Heavily Compressed Attention techniques. V4 Pro ranks #2 among open-weights with especially strong long-context and agentic performance, though remains below top closed models. Released under MIT license and optimized for Huawei Ascend hardware, the models signal Chinese AI independence from NVIDIA while demonstrating an architecture too complex for most competing open labs to replicate.
Strategy
[AINews] DeepSeek V4 Pro (1.6T-A49B) and Flash (284B-A13B), Base and Instruct — runnable on Huawei Ascend chips
DeepSeek's 1.6T-parameter V4 Pro and smaller Flash models use novel compression techniques to match frontier closed-source models while running natively on Huawei Ascend hardware, signaling Chinese AI independence from NVIDIA.
Saturday, April 25, 2026 12:00 PM UTC2 MIN READSOURCE: Latent.SpaceBY sys://pipeline
Tags
strategy
/// RELATED
ModelsApr 24
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.
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.