NVIDIA ReplaySSM Rethinks State Storage for Faster Hybrid-Model Decoding
NVIDIA published ReplaySSM, a technique for faster decoding in hybrid models that interleave state-space model layers with attention layers. The post names NVIDIA Nemotron 3, Qwen3.5 through Qwen3.8, Kimi-K3, and GLM-5.3-Flash as models built this way, using Mamba-2, Gated DeltaNet, or Kimi Delta Attention as the SSM layer.
Sources and evidence
Summary last validated Oct 12, 2026
Reader actions
Report an issue
Use this for an incorrect summary, wrong source, duplicate story, or wrong category. Submissions are private and do not change the story automatically.
Related coverage
- infrastructure
Barcelona Supercomputing Center Opens AI Factory Accelerator to Spanish Startups
The Barcelona Supercomputing Center's AI Factory Accelerator Programme is now in full operation after a pilot phase, giving startups access to sovereign compute, technical support and training across industries. Many participating companies are also members of the NVIDIA Inception program for startups, and as of today some gain expanded access.
- industry
NVIDIA and Germany Discuss Industrial AI Leadership at Berlin Event
At Berlin's Kraftwerk venue, German Chancellor Friedrich Merz and NVIDIA CEO Jensen Huang discussed Germany's AI roadmap, focusing on industrial AI applications. Huang traced NVIDIA's long relationship with German industry, from gaming systems and workstations to automotive. The event was framed around Germany's ambition to lead the world in industrial AI, with NVIDIA positioned as a partner.
- research
MIT and Boston Children's Hospital develop symmetry-aware AI for fetal ultrasound
Researchers at MIT and Boston Children's Hospital developed a model that teaches AI about symmetry to adjust fetal ultrasound scans on the fly, tracking fetal movement during imaging. The approach aims to make prenatal diagnosis more reliable and could spare families repeat hospital visits when scans are inconclusive.
- open source
NVIDIA Shows Mamba-2 Forward Pass Kernels in CUDA Tile Python DSL
NVIDIA published a developer blog demonstrating high-performance Mamba-2 forward pass kernels written in CUDA Tile (cuTile), its Python-based DSL and parallel programming model for NVIDIA GPUs. cuTile abstracts thread- and warp-level scheduling behind tile-centric primitives, letting developers write GPU kernels in Python while compiler optimizations handle low-level hardware details automatically.
- research
Meta details GenRecs tokenization for generative ads recommendation
Meta published an engineering post on GenRecs Tokenization, the data representation layer behind its Generative Recommenders for ads. Meta says the system shifts from scoring every ad to reasoning over ad content and user preferences using LLM world knowledge, with a stable model vocabulary intended to make ad matching more precise and compound gains for advertisers.