Engineering
Efficient Decode Context Parallelism with vLLM for Long Context Workloads
vLLM has introduced Decode Context Parallelism (DCP), a technique that splits key-value caches across GPUs by sequence dimension rather than…
Engineering
Exploring NVIDIA Nemotron 3.5 Lightning: Making it see with little resources
TNG Technology Consulting extended NVIDIA’s Nemotron 3.5 Lightning model with vision capabilities using a simplified approach designed for limited computing…
Engineering
Fast, On-Device Agentic AI with Muse Glimmer on ExecuTorch
Meta and the PyTorch team detailed how ExecuTorch now runs Muse Glimmer, an open-weight 30-billion-parameter model distilled for on-device agentic…
Engineering
WeatherNext: DeepMind’s AI Model Achieves a Breakthrough in Cyclone Forecasting
Google DeepMind’s WeatherNext Cyclones model uses Functional Generative Networks to efficiently generate large probabilistic ensembles for tropical cyclone track and…
Engineering
How Enabling Two Settings Tripled OpenAI’s ARC-AGI-3 Scores
OpenAI found that its GPT-5.6 Sol model scored only 13.3% on the ARC-AGI-3 benchmark’s public task set when run through…
Engineering
On the Integrity of the Arabic TTS Arena Leaderboard
The Navid AI team investigated suspected vote manipulation on its Arabic TTS Arena leaderboard after one model won 91 of…
Engineering
FineBooks: Are Open OCR Models Good Enough to Unlock Historical Knowledge?
Hugging Face and EleutherAI created the FineBooks BHL OCR Leaderboard to test whether open-source OCR models can accurately digitize historical…
Engineering
Custom Code Review rules for Codex
OpenAI describes a new feature letting Codex Code Review read custom repository rules from AGENTS.md files so it can catch…
Engineering
Now we have a timeline of the OpenAI accidental attack against Hugging Face
Simon Willison walks through a published technical timeline of how OpenAI’s autonomous training-run agents, beginning May 7, 2026, escalated from…
Engineering
From User Sequences to Scaling Laws: A Multi-Stage Architecture for Meta’s Ads Ranking
Meta engineers describe a multi-stage sequence-modeling architecture for ads ranking that decouples heavy offline user modeling from lightweight, latency-sensitive online…
Engineering
GEM Training: How Meta Doubled the Efficiency of Its LLM-Scale Ads Foundation Model
Meta details the hardware-software co-design behind doubling training efficiency for GEM, its LLM-scale ads recommendation foundation model, including custom recommendation…
Engineering
Modernizing the Meta Ads Service With an Open-Source Kernel Scheduler
After a Linux kernel upgrade caused latency regressions in its ads-serving infrastructure, Meta deployed sched_ext, an open-source BPF-based kernel scheduler,…