🚀 Now serving MiniMax-M3 for efficient inference →

📊 Delivering 31% more TPS than the next-fastest OSS engine for production coding agent workloads →

💬 How Together built the world's fastest speech-to-text stack →

🇫🇷 Join us at RAISE 2026 in Paris →

Explore Research

Research blog

All
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
Architecture
Mamba-3

    Aakash Lahoti* (CMU), Kevin Y. Li* (CMU), Berlin Chen* (Princeton), Caitlin Wang* (Princeton), Aviv Bick (CMU), J. Zico Kolter (CMU), Tri Dao (Princeton, Together AI), Albert Gu (CMU, Cartesia A)

    Abstract purple and red shapes background with the word 'Mamba-3' and a small label 'RESEARCH'.
    Kernels
    Key research and product announcements at the AI Native Conf

      Kernels
      FlashAttention-4: Algorithm and Kernel Pipelining Co-Design for Asymmetric Hardware Scaling

        Ted Zadouri (Princeton University, Together AI), Markus Hoehnerbach (Meta), Jay Shah (Colfax Research), Timmy Liu (NVIDIA), Vijay Thakkar (Meta, Georgia Tech), Tri Dao (Princeton University, Together AI)

        Abstract colorful curved shapes with the text FlashAttention-4 and a small label reading RESEARCH.
        Inference
        Cache-aware prefill–decode disaggregation (CPD) for up to 40% faster long-context LLM serving

          Jiejing Zhang, Yubo Wang, Yinghui Liu, Mourya Vangala Srinivasa, Chenxi Li, Jue Wang, Yineng Zhang, Shuaiwen Leon Song, Ce Zhang

          Bar chart showing CPD outperforming baseline in queries per second for 1D vs 2P1D and 2D vs 2P2D tests.

          No search result

          Try expanding your search or changing the filters.