New model releases target Arabic, multimodal, and edge deployment

Jais 3 combines Arabic LLMs from phone-scale to 375B with speech recognition and tool workflows; its developers report speech benchmark gains that need independent confirmation. NVIDIA announced a 120B multimodal model and two edge models, with performance claims not yet independently verified. MOSS-Image adds open-weight structured-layout generation under Apache 2.0, while Reflection’s Apache 2.0 Beam 501B reports 81.6% on SWE-bench Verified, pending independent testing.