AI Intel

Gemma 4 E2B has demonstrated in-browser inference at 255 tokens/second on an M4 Max, utilizing WebGPU kernels optimized with assistance from Fable 5.

New research on arXiv presents "When Rules Learn," a self-evolving agent designed to improve legal case retrieval by addressing the complexities of l…

PhoneHarness is a new framework that enables phone-use agents to complete mobile workflows through a mix of GUI, CLI, and tool actions, addressing li…

GLM-5.2 has achieved a significant milestone by becoming the first open-weights model to score over 80% on Terminal-Bench, demonstrating frontier-lev…

Ollama's v0.30.9 update introduces support for the Cohere2Moe architecture and resolves issues with LFM2 parser/render and coding agent outputs.

Claude Code's v2.1.178 update introduces advanced Tool(param:value) syntax for permission rules, allowing more granular control over tool input param…

Locus Founder is a new AI agent product that allows users to build and operate their businesses by simply texting the AI.

Deep Work Plan highlights the critical importance of providing AI agents with context and structured plans to enhance their performance, beyond just…

Tyto, developed by ai-coustics, offers audio insights capable of predicting the performance of voice AI systems.

vLLM has released version v0.23.0, incorporating 408 commits from 200 contributors, though Minimax M3 is not yet supported.