MiniMax M3 combines coding, agentic capabilities, vision, video understanding and a functional one-million-token context window.
2
MiniMax Sparse Attention uses an index branch to select important parts of a context before a sparse attention branch processes selected blocks.
3
MiniMax trains multimodal capability from the first step because adding vision after text training can harm text performance and produce weaker vision understanding.
Summary
Olive Song presents MiniMax M3 as a model built for agents that need to retain long conversations, tool responses and information from complex environments. M3 has around 400 billion total parameters, 20 billion activated parameters, vision and video understanding, and a functional one-million-token context window. Its MiniMax Sparse Attention architecture first identifies important blocks, then applies attention to those blocks. Song says the team trains multimodal capability from the beginning rather than adding vision after text pretraining, because later integration can harm text performance and fail to converge well on vision. She describes MiniMax's open research process, where people can propose projects after a release and contribute work that enters later training. The team also uses research harnesses to automate data creation and other workflows. M3 is already being used to help build M3.1, while Song expects multi-agent systems and model routing to expand what models can do.
M3 joins long context with coding and multimodal understanding
Olive Song describes M3 as a model with around 400 billion total parameters and 20 billion activated parameters. It handles coding and understands images and videos, which Song says is unusual among open-source models. It also has a one-million-token context window through MiniMax Sparse Attention. MiniMax brought these capabilities together because the team expects coding, agentic work, long context and multimodal understanding to matter in future AI applications. Thomas Wolf adds that M3 was the only top-five open-source model he knew of that was multimodal at the time of the discussion.
Agents need context that includes their whole working environment
Song says MiniMax had already tested models that could handle 10 million tokens, but those models were not agentic. They could take in a book and review it, for example. The team then saw that longer context mattered more when a model had to interact with users and an environment. An agent accumulates tool responses and multiple rounds of interaction, so a short context cannot hold enough information for complex tasks. M3 therefore brought back a long context as part of a model designed for agentic use rather than document processing alone.
Sparse attention reduces the work needed for long context
MiniMax Sparse Attention has two parts. An index branch selects what matters most in the context at a higher level. A sparse attention branch then performs the attention calculation on the selected blocks. Song says the architecture has a simple design that can scale context length and model size. Wolf connects this to earlier attempts to make attention cheaper, including linear attention and the later focus on efficient implementations. Song expects further work on architecture and inference optimization, especially for tasks that need strong capabilities while remaining efficient.
An intern designed the architecture in an open research environment
Song says an intern on the MiniMax team designed the sparse attention architecture. She contrasts this with labs where interns may not have access to data or substantial work. At MiniMax, the team builds shared infrastructure so people can experiment with the model, create their own evaluations and identify weaknesses after a release. Anyone can propose a project, and interested colleagues can join for weeks or months. Architecture work can take longer because it involves research, experiments and repeating pretraining evaluations. Successful work enters the final training process and is shipped in a later model.
Native multimodality starts during the first training step
MiniMax calls its approach native multimodality. Song says that adding vision after text pretraining can harm text performance, while vision understanding may not converge well because the model has already moved toward text understanding. Continuing pretraining halfway through is also sensitive to the architecture, data mixture and learning rate, which makes results hard to control and scale. MiniMax instead trains with text, images and videos from the first step. The team worked on the vision transformer, data cleaning and masking, interleaved natural data, and reward modeling. Song says these changes prevented the model from collapsing during training.
MiniMax expects larger models and already has broad product usage
Song says M3 has 428 billion parameters in total and 23 billion active parameters, using a more precise figure later in the conversation. She expects MiniMax to go beyond a trillion parameters because some tasks need more capacity than a smaller model can provide. She also explains that MiniMax built user-facing apps so people could experience its models without using an API. Those apps reach more than 300 million people across around 200 countries, and Song says more than a million companies use them. The products existed alongside the research effort rather than being created only after the models were developed.
Song says the research team plans to keep open-sourcing its models. The community sends feedback about model performance and pull requests for released work, and MiniMax uses that input in later versions. She specifically asks users to report multimodal failures and requests for features such as thinking effort. Song acknowledges that the combined multimodal system still has flaws and says the team wants to improve it in future models. The exchange presents open source as part of the model development process, because external use reveals problems and suggests changes the internal team may not have found.
Long videos and research harnesses extend agent workflows
Song says multimodal models can let an agent read a PowerPoint, understand an unstructured report or watch a long video before using tools. Thomas Wolf gives the example of an agent watching a coding tutorial and learning how to use the described tools, and Song agrees that this is possible. Inside MiniMax, the team has built research harnesses that automate many workflows. Song mentions model-generated data and models operating on other models as examples of the wider trend. MiniMax uses these capabilities in daily research work, and M3 is already helping the team build M3.1.
"Longer context actually unlocks a lot of capabilities especially when interacting with users and now when the agent is interacting with the whole environment and getting all the tool responses."Olive Song04:31
Who should watch
You are choosing a context strategy for coding agents that must retain tool calls and long interaction histories.
You are training a multimodal model and want to understand the problems with adding vision after text pretraining.
You work on model research infrastructure and want examples of open project selection, community feedback and automated research harnesses.
