Hugging Face AI Updates: May 28, 2026

1. ITBench-AA Shows Frontier Models Still Below 50% on Enterprise IT Tasks

Hugging Face. Artificial Analysis and IBM Research published ITBench-AA, a benchmark for agentic enterprise IT tasks (incident response, change management, configuration troubleshooting), and frontier models all land below 50%. The benchmark is interesting because it scores on action correctness end-to-end rather than just intermediate reasoning, which is where most existing agent benchmarks quietly grade on a curve. Source

2. Reachy Mini Goes Fully Local

Hugging Face. The Reachy Mini robotics platform can now run conversation and motor-control models entirely on-device, without cloud round-trips. The post walks through the quantization and routing choices that make a small-LM speech-and-action loop tractable on the embedded compute Reachy Mini ships with, which is the practical question every consumer-robot startup is facing right now. Source

3. Delta Weight Sync Lands in TRL for Trillion-Parameter Training

Hugging Face. A new delta-weight-sync mechanism in TRL lets you train models with up to a trillion parameters while only persisting weight deltas back to a Hub bucket between steps, slashing storage and bandwidth costs versus full-checkpoint snapshots. This is the kind of unsexy infra change that quietly determines whether 1T+ parameter open-source training is actually affordable for outside groups. Source