Overview
MiniMax M3 and Meta's Muse Spark Open are both large multimodal models aimed at agentic and coding work that you can run on your own hardware. Muse Spark Open's weights are available now. MiniMax describes M3 as open-weight and says the full open-source release on Hugging Face and GitHub is coming soon.
MiniMax M3
Built on MiniMax Sparse Attention. The API supports up to 1M tokens of context, with a guaranteed minimum of 512K.Natively multimodal, with multimodal training from the start of pretraining.Vendor-reported results: BrowseComp 83.5 against 79.3 for Claude Opus 4.7, and PostTrainBench 37.1.Available through the MiniMax API (model MiniMax-M3) and MiniMax Code. MiniMax does not state a parameter count.Meta Muse Spark Open
The open-weight release of Muse Spark: an 816B-parameter mixture-of-experts model with 42B active parameters.Accepts text, image, video and audio input, with text output.Weights are available in BF16 and NVFP4 for SGLang, vLLM, TensorRT-LLM or llama.cpp, or you can use a hosted inference provider. Meta supports fine-tuning with SFT and RL.Meta's models page does not state a context length for Spark Open. The hosted Muse Spark versions have a 1,048,576-token window.How to choose
If you need weights today, Muse Spark Open is downloadable now. MiniMax M3's full weights are announced but not yet released.If you want a hosted API with a published long-context guarantee, MiniMax states 1M tokens with a 512K minimum.Check each licence on the vendor pages before commercial use.No hands-on testing was conducted for this comparison. Every figure above comes from the official pages listed under Research sources, as published on or before October 22, 2026. Where a vendor has not published a number, we say so instead of estimating it.