Vercel's AI Gateway is offering Baseten's two decision models free of charge until October 27. Mammoth-1 has 27 billion parameters and Elephant-1 has 9 billion; both classify text, route requests and score content against custom criteria, with a 16k-token context window.
NVIDIA has presented Cosmos3-H-Dreams on the Hugging Face blog, describing it as a way to recreate clinical surgery through generative simulation.
Mistral AI has released Voxtral 3, a family of models for real-time voice agents that bundles speech-to-text, text-to-speech and a voice-agent SDK. It covers more than 50 languages, lets agents call external tools mid-conversation, and can be self-hosted or used through EU-hosted APIs for data control.
Interconnects' regular roundup of open models covers DeepSeek V4.1 Flash 1017, new American-made models and fresh regulatory controls in the US and China. The author frames it as Western open models striking back.
ElevenLabs unveiled two speech-to-speech models, Mini and Max, designed as complete voice agents. They are available now and compatible with OpenAI's Realtime API.
Scaled Cognition, which builds models for customer service agents, introduced Voice, a text-to-speech model. The company says it pronounces account numbers and addresses more accurately and costs less than rivals; it is available via API or self-hosted.