Alibaba has launched its new AI model, Qwen3-235B-A22B-Instruct-2507, on the Hugging Face platform, marking an advancement over the previous Qwen 3 model in reasoning, accuracy, and multilingual understanding. This upgraded model, which employs a mixture-of-experts (MoE) architecture, is recognized as the most intelligent non-reasoning model, surpassing competitors such as Kimi K2 and Claude 4 Opus on the Artificial Analysis Intelligence Index. The release includes versions like Qwen3-Coder and Qwen3-2507, now available on Weights & Biases Inference with competitive pricing and incentives such as $20 credits for users. The model supports day-zero inference and benefits from infrastructure support by CoreWeave, enabling cost-effective deployment on H200 hardware. Separately, Boson AI introduced Higgs Audio V2, an open unified text-to-speech model with voice cloning capabilities, trained on 10 million hours of diverse audio data. Built on Llama 3.2 3B, Higgs Audio V2 operates in real-time, including on edge devices, and outperforms GPT-4o-mini-tts and ElevenLabs v2 in prosody.