Hugging Face AI Updates: June 6, 2026

1. NVIDIA Releases Nemotron 3.5 Content Safety, a 4B Multimodal Guardrail Model

Hugging Face. NVIDIA published Nemotron 3.5 Content Safety on the Hugging Face Hub, a 4-billion-parameter model that evaluates user prompts, images, and assistant responses together to catch policy violations arising from text-image interactions. The model supports 12 languages explicitly and over 140 through zero-shot transfer, accepts domain-specific safety policies at inference time, and offers an optional THINK mode that produces auditable step-by-step justifications for its verdicts. NVIDIA reports roughly 85 percent accuracy across multimodal and multilingual benchmarks at about 3x lower latency than comparable alternatives. Source

2. Hugging Face Redesigns the hf CLI for Coding Agents

Hugging Face. Hugging Face rebuilt its hf command-line tool to detect agent usage through environment variables and adjust output accordingly, giving humans colorized tables while agents receive compact TSV data without formatting overhead. In benchmarks the CLI reached a 94 percent task success rate versus 84 percent for agents using curl or the Python SDK directly on Claude Code, while consuming far fewer tokens, with the largest gains on multi-step tasks. Added features include next-command hints, safe retries, and an auto-generated skill reference that cuts tool calls by roughly 30 percent. Source

3. ServiceNow Publishes EVA-Bench Data 2.0 for Enterprise Voice Agents

Hugging Face. The ServiceNow AI team released EVA-Bench Data 2.0, a benchmark for evaluating voice agents across three enterprise domains spanning 121 tools and 213 scenarios, roughly a fourfold expansion over the original. The domains cover airline customer service, enterprise IT service management, and healthcare HR service delivery, with scenarios built around single deterministic resolution paths for reproducibility. The dataset is released under an MIT license and available through the Hugging Face datasets library. Source