Hugging Face Blog
·
1 month ago
NVIDIA released Nemotron 3.5 Content Safety, a 4B-parameter model that evaluates text, images, and responses simultaneously to detect policy violations that emerge only from their interaction. The model achieves 85% average accuracy across multilingual and multimodal safety benchmarks while maintaining low latency on 8GB+ VRAM GPUs. Enterprises can now enforce custom domain-specific policies at inference time and receive auditable reasoning traces explaining each safety decision, rather than relying on fixed universal taxonomies.
Ben's Bites
·
1 month ago
A newsletter discusses progress on building an agents manual and reflects on how creating tools enables further tool-building, followed by a roundup of recent AI releases including new open models, agent products, and agent-focused tooling across multiple domains.
IBM Research
·
1 month ago
IBM released Granite Libraries, a set of adapter tools that break down large language models into modular, task-specific components similar to software building blocks, allowing developers to customize AI behavior without retraining entire models. The Granite 4.1 8B model with requirement-check adapters achieved 84% balanced accuracy on the IFEval instruction-following benchmark, compared to 51% for the base model with prompting alone. This modularity enables multiple teams to incrementally add capabilities, reduce inference costs, and make AI systems more predictable and maintainable for enterprise deployment.
Hugging Face Blog
·
1 month ago
Hugging Face redesigned its hf command-line interface to work efficiently with coding agents like Claude Code and Codex, which now drive a significant portion of Hub traffic. Testing showed that agents using the hf CLI consumed 1.3 to 1.8 times fewer tokens than those using curl or the Python SDK directly on complex multi-step tasks, with the CLI achieving 93-94% success rates versus 84-92% without it. The optimized CLI auto-detects agent usage, outputs dense TSV format instead of human-readable tables, and provides chainable next-step hints to reduce the number of commands agents need to execute.
