New arXiv preprint introduces Page-EntroKV, a hardware-aligned eviction method using Renyi-2 entropy to pool query heads under GQA. It eliminates union overhead, prevents sink…
An arXiv preprint from October 2 2026 investigates whether forcing reasoning in the same language as the query and retrieved documents improves accuracy in retrieval-augmented…
arXiv preprint proposes a lightweight Zero-and-First-Order framework that uses a first-order optimizer for direction and only two extra objective evaluations for…
arXiv preprint introduces TACO, a ternary one-sparse optimizer that follows an operator-norm steepest-descent direction by selecting the sign of the largest-magnitude entry per…
Allen Institute for AI releases AstaBrief 8B, a specialized open-weights model that produces cited scientific reports from queries and literature excerpts. It delivers reports…
Meta's new 30B-parameter multimodal model, distilled from Muse and released under Apache 2.0, targets local deployment for privacy-focused coding, document analysis and…
Meta Superintelligence Labs released its first open model, the 30B-parameter multimodal Muse Glimmer, on August 10 2026. Optimized for local agent workloads with 128K+ context…
NVIDIA's Nemotron 3.5 Lightning, a 30B-parameter Mixture-of-Experts model with 3B active parameters per token, is now available via Ollama for fully local inference. Optimized…
New research quantifies how leading ASR systems reproduce erroneous reference transcripts, recover silenced numbers, and switch orthography to match public benchmarks like…
New @huggingface/kernels JavaScript library and 207 WebGPU kernels published on the Hub provide optimized, inspectable operations for browser-based local AI. Fleet crowdsources…