Show HN: Getting GLM 5.2 running on my slow computer
colibrì enables running the GLM-5.2 (744B-parameter MoE) model on consumer hardware with ~25 GB of RAM, utilizing a unique architecture that streams experts from disk, allowing for efficient memory usage and processing.
Inference Optimization for MiMo v2.5: Pushing Hybrid SWA Efficiency to the Limit
The MiMo-V2.5 series leverages Hybrid Sliding Window Attention (SWA) to achieve a 7× reduction in both compute and KVCache storage costs, enhancing efficiency for long-context and multimodal tasks.
Computation as a universal and fundamental concept
Computation's limits were first established by Alan Turing, who demonstrated that certain problems, like the halting problem, are unsolvable by any algorithm, highlighting the inherent boundaries of computational power.
Silent speech with ultrasound
Model predicts speech from ultrasound recordings of the tongue, achieving a 15.6% word error rate on open-vocabulary speech, which is competitive with lip-reading methods trained on larger datasets.
Nvidia, CoreWeave, and Nebius: Inside the Circular Financing of the GPU Boom
Neoclouds like CoreWeave and Nebius are capitalizing on the GPU boom by securing massive contracts, with commitments from Microsoft and Meta exceeding $120 billion, significantly outpacing their current revenues.
Show HN: Reame – a CPU inference server that gets faster as it runs
Reame is a pioneering LLM inference server that optimizes for low-cost CPU hardware, achieving 100% accuracy on long-context tasks with minimal resource expenditure, making it ideal for repetitive AI workloads on existing infrastructure.
VultronRetriever family of models released on HuggingFace!
The VultronRetriever family of models excels in performance, with VultronRetrieverPrime-8B ranking as the global #1 on the MTEB Leaderboard, showcasing a 16x smaller index storage and 12x higher throughput than previous leaders.
Towards Free Normalization: Fusing Normalization into GEMM and Attention Kernels
Novel kernel fusion techniques for normalization operations like LayerNorm and RMSNorm can achieve up to 90% latency reduction by integrating them with GEMM and attention kernels, significantly enhancing performance in deep learning models.