LiquidAI's LFM2.5-2.6B runs agentic workflows on-device with 128K context, tool calling, and an OpenAI-compatible endpoint. No cloud inference bill.
Hugging Face revamped Kernels with cryptographic signing, trusted publisher verification, and first-class Hub integration. Here's what changed.
Photoroom's PRX Part 4 details their training data strategy — sourcing, curation, synthetic data, and open dataset releases on HuggingFace.
Hugging Face and Cerebras built a real-time voice AI pipeline using Google's Gemma 4 model, cutting latency to conversational speeds with open-source components.