Researchers from UC Berkeley's UCCL project released mKernel, a library of persistent CUDA kernels that fuse intra-node NVLink communication, inter-node RDMA, and compute into a single kernel to addre…
Hexo Labs released SIA (Self-Improving AI) as an open-source framework under an MIT license this week, enabling an AI agent to edit both its scaffold and model weights within a single self-improving l…
Liquid AI released LFM2.5-8B-A1B, an on-device Mixture-of-Experts model with 8.3 billion total parameters and 1.5 billion active parameters per token, designed for tool calling and reasoning on consum…
Anthropic launched Claude Opus 4.8 alongside two Claude Code updates in research preview, including dynamic workflows that orchestrate up to 1,000 subagents in parallel and a cheaper fast mode deliver…
Perplexity AI open-sourced a Unigram tokenizer reimplemented in Rust that achieves 5x lower p50 latency than the Hugging Face tokenizers crate at production input lengths. The new tokenizer, available…
A new coding tutorial demonstrates how to build a complete pgvector-powered vector search system inside Google Colab, supporting semantic, hybrid, sparse, and quantized search methods. The guide walks…
Researchers from Sakana AI and the University of Tokyo developed DiffusionBlocks, a block-wise training framework that converts residual networks into independently trainable denoising modules. The me…
NVIDIA researchers released Polar, a rollout framework that trains language agents through reinforcement learning without altering their agent harnesses. Using GRPO on a Qwen3.5-4B base model, Polar i…
The EAGLE Team, vLLM Team, and TorchSpec Team released EAGLE 3.1, a speculative decoding algorithm that addresses attention drift in large language model inference. The update introduces FC normalizat…
Researchers from the National University of Singapore, MIT CSAIL, A*STAR, and SMART have developed MEMO, a framework that trains a separate, small language model to internalize new knowledge from docu…