Novel attention architecture achieving 354M parameters with long-context VRAM savings via structural sparsity and custom Triton kernels. Open weights released.
DVD-JEPA is a fully-reproducible JEPA world model released as open-source, addressing reproducibility in world modeling research.
Alibaba's Qwen-Agent framework has accumulated 16,587 stars, indicating substantial adoption for building LLM-based agents.
DeepSeek-R1 repository shows significant activity with 91,986 stars. Represents latest iteration of Chinese open-source reasoning model gaining substantial adoption.