Hugging Face AI Updates: July 16, 2026

1. Thinking Machines Releases Inkling, a Trillion-Parameter Open Multimodal Model

Hugging Face. Thinking Machines published Inkling on the Hugging Face Hub, an open-source multimodal model of roughly 1 trillion parameters with a 1 million token context window that processes text, images, and audio. Built on a decoder-only Mixture-of-Experts architecture and trained on 45 trillion tokens across text, images, audio, and video, it targets multimodal reasoning and agentic coding and includes speculative decoding. It ships in full precision (2TB VRAM) and quantized (600GB VRAM) formats with support across transformers, SGLang, and vLLM. Source

2. Ai2 Shares Lessons From Building Shippy, a Maritime Domain Agent

Hugging Face. Ai2’s Skylight team published a technical post on Shippy, an AI agent for maritime domain awareness that answers operational questions about vessel activity, fishing behavior, and ocean protection. The post argues for decomposing agents into modular parts, a system prompt “soul,” reusable skills defined in markdown, and swappable configuration, and for wrapping complex APIs in deterministic CLI tools to avoid subtle bugs. It also describes building custom evaluation frameworks that test the full agent against live data rather than static benchmarks, with subject-matter experts defining success. Source

3. IBM Research Details the Hidden Complexity of Model Routing

Hugging Face. IBM Research published a post arguing that effective model routing in agentic systems requires optimizing across many system factors rather than selecting models by task difficulty alone. The authors, including Yara Rizk, Eyal Shnarch, Jason Tsay, and Merve Unuvar, note that real costs depend on caching behavior and infrastructure rather than list pricing. They conclude that routing decisions must balance cost, latency, compliance, and reliability at once rather than being treated as a simple classification problem. Source

4. Hume AI Launches Real World VoiceEQ Benchmark for Voice AI

Hugging Face. Hume AI introduced Real World VoiceEQ, a benchmark that evaluates whether voice AI systems can recognize, produce, and respond to the acoustic information transcripts omit, such as tone, emotion, speaker identity, and background context. The benchmark assesses more than 40 voice models across over 15 dimensions and 60 metrics. Its scoring draws on more than 1 million human ratings. Source