MiniMax-M3 sits at the frontier of MiniMax's M-series as a Mixture-of-Experts model with 428 billion total parameters and roughly 23 billion active per forward pass, paired with native multimodal input across text, image, and video. Its long-context scaling is anchored by MiniMax Sparse Attention (MSA), an architecture explicitly designed to make very large that quick-info value windows practical rather than merely nominal, with a published the cataloged API limit capacity and vendor-reported 500K rollout that quick-info value on some endpoints at launch. Open weights were released to the official Hugging Face repository shortly after launch, reinforcing its positioning as a fully open-weight frontier option rather than a closed API product.
Independent benchmark and adoption signals point to M3 being a strong fit for agentic coding workflows, with vendor-cited results placing it at 59.0% on SWE-Bench Pro and significant real-world usage through the OpenCode Go provider, where it ranks among the top models by token consumption. It shares that provider's catalog with other open coding-oriented models such as MiniMax M2.7, GLM variants, Kimi K-series, and Qwen3 models, but its combination of open weights, sparse-attention long that quick-info value, and multimodality gives it a distinctive profile for teams that want frontier agentic behavior without proprietary lock-in. Practically, it is well suited to long-horizon repository reasoning, tool-driven coding agents, and tasks that benefit from ingesting video or image that quick-info value alongside source code.
