Meta released a 30B multimodal model built for local agent workloads — beats Gemma4 and holds its own against Qwen3.6 on agentic benchmarks, and fits on a single consumer GPU. Here's what's actually new, and what it'd take for it to land on WEC.