Multimodal Models

MiniMax M3 is a frontier-tier open-weight multimodal agent model with 428B total parameters (23B activated) and a 1M token context window. Released May 31, 2026, it features MiniMax Sparse Attention (MSA) for 9x prefill and 15x decode speedup over its predecessor, and is specialized for coding, agentic tasks, and complex reasoning.

NVIDIA Cosmos 3 is an open frontier foundation model for physical AI, built on a Mixture-of-Transformers (MoT) architecture. It unifies text, image, video, ambient sound, and action generation within a single omnimodal system designed for robotics, autonomous vehicles, and physical environment simulation. Available in Nano (16B) and Super (64B) variants.

Gemini 4 Argon is Google's frontier multimodal reasoning model, made publicly available through the Gemini API on October 13, 2026. It features a 1 million token context window and is designed for long-horizon, complex agentic workflows. Introductory API pricing of $2 per million input tokens and $10 per million output tokens applies through November 30, 2026.

Meta Muse Spark Open is an open-weight multimodal model released by Meta AI on October 15, 2026, targeting coding and agentic tasks. It features a sparse mixture-of-experts architecture with 816 billion total parameters and 42 billion active parameters per token, with a long-context window designed for multimodal reasoning and agent workflows.