November 2026 Summaries

No summary generated yet.

DeepInfra has made Xiaomi’s MiMo-V2.6-Pro available through its OpenAI-compatible inference API, positioning the MIT-licensed model for long-running coding, security, and multimodal agent workflows. The sparse mixture-of-experts model has 1.02 trillion total parameters but activates 42 billion per pass, supports text, image, video, and audio, and offers a one-million-token context window. Xiaomi attributes major gains over MiMo-V2.5-Pro to a mixed-domain reinforcement-learning approach called “You Only RL Once,” along with agentic grading and speculative decoding; its reported results show especially large improvements on DeepSWE, AutomationBench, CyberGym, and other agent-oriented evaluations, while results remain behind some competing models on ExploitBench and Terminal Bench. DeepInfra offers standard and faster UltraSpeed serving modes, charges $0.435 per million uncached input tokens and $0.87 per million output tokens, and highlights prompt caching, private deployments, zero data retention, and SOC 2 and ISO 27001 certifications for production users.