UC San Diego's BeNice Slashes AI Memory Use by 66% in Looped Models
A new two-stage training recipe lets looped language models reuse a single KV-cache across every recurrence, cutting HBM use up to 66%.
20 stories
A new two-stage training recipe lets looped language models reuse a single KV-cache across every recurrence, cutting HBM use up to 66%.
IFM open-sources the agentic reinforcement learning stack behind K2 Horizon, running 8,192 concurrent sandboxes and 1.3M rollouts across seven days.
Liquid AI's new RIPL inference engine runs small Liquid Foundation Models in under 50ms end-to-end, targeting search, ads, fraud, and payments.
A 4-bit quantized build of Qwen2.5-3B-Instruct runs chat, coding and long-context tasks on roughly 2.5 GB of VRAM.
A Claude Code fix passed 9 required tests and still returned the wrong title. One closed question after the tests picked the suite that caught it
Ai2 released OlmoEarth v1.3 with precomputed global satellite embeddings at 10m resolution and a browser tool that trains classifiers without a GPU.
METR and Redwood Research embedded at OpenAI for three weeks and found multiple exploits that let agents evade internal misalignment monitors.
Prompt filters and borrowed human tokens don't hold up once agents reach production. Ory Agent Security splits the problem in two: authorize every tool call inside the agent loop, and give every deployed runtime its own machine identity.
Scenario open-sourced a Blender 5.0 extension packing 500+ generative models for images, video, 3D, audio and PBR, plus a local MCP server for agents.
Google DeepMind's ForTy classifies every 10x10m patch of Earth into six forest types, revealing not just where trees grow but what kind.
Kyutai's CADRE replaces the laggy video autoencoders inherited from text-to-video models, letting interactive world models respond in real time on one GPU.
Wes Bos released an open-source toolkit that turns iPhone LiDAR scans into editable Three.js scenes through an AI coding agent workflow.
AgnesAI Labs open-sourced a pluggable TypeScript runtime that lets agents share tools, tasks, and interfaces across CLI, Web, and physical devices.
A new multilingual benchmark exposes how embedding models and LLMs fail at fact-checking across 10 African languages, and shows what fixes work.
A community fine-tune of Qwen3.8-27B trades a sliver of raw accuracy for shorter reasoning traces, faster decoding, and cleaner Python output in long agent loops.
Tencent and Fudan built 16,700 verifiable training environments from real files across 40 professional domains, lifting a 35B agent's GDPval score by 28 points.
A DSH plugin ships frontier models like DeepSeek V4.1 Flash and Kimi K3 inside the official agent harness with no login, key, or quota.
Recursive's first model is a 40B mixture-of-experts coder that activates just 2.5B parameters, runs on a MacBook, and was largely designed by an automated research agent.
A new point cloud backbone uses convolutions early and attention late, cutting parameters 3.6x while matching Point Transformer V3 accuracy.
Nous Research ships an official Meshy plugin for Hermes Agent, bringing text-to-3D, rigging, animation and 3D print prep straight into chat.