DeepSeek-V4 hits Hugging Face with 1.6T MoE, 1M context

DeepSeek-V4 is a direct challenge to the top-tier closed-source models, doubling down on the "efficient MoE" architecture that made V3 a developer favorite.

• –1M context window becomes the new baseline for foundation models, supported by novel compressed attention architectures that reduce memory overhead.
• –V4-Pro (1.6T) targets elite-level coding and reasoning performance, reportedly rivaling Claude 4 and GPT-5 class models in technical benchmarks.
• –V4-Flash (284B total, 13B active) is a massive efficiency play, likely to dominate the high-throughput, long-context agentic market.
• –Engram Conditional Memory and Manifold-Constrained Hyper-Connections (mHC) signal a shift from simple scaling to deep architectural refinement for signal stability.
• –MIT licensing and aggressive pricing continue to erode the competitive moat of closed-source API ecosystems.