MiniMax-M3

MiniMax-M3 is a multimodal model for long-context coding, perception, and agent planning. It is the latest release in the MiniMax line, succeeding MiniMax-M2.7. MiniMax-M3 has a 1,000,000 token context window and a 512,000 token output limit. It accepts image, text, and video inputs and returns text. It supports reasoning mode and tool calling. Weights are open. Price starts at $0.20 per million input tokens and $0.90 per million output tokens, with cached input at $0.045 per million tokens.

Against MiniMax-M2.7: window 205K to 1M, input $0.08 to $0.20, output $0.32 to $0.90, cached input $0.017 to $0.045.

Range across the hosts that serve it: up to $0.72 per 1M input tokens.

• ·1M token context window
• ·512k token output limit
• ·Accepts image, text, video
• ·Reasoning mode and tool calling
• ·Open weights

• Reach for it for long-context coding tasks
• Reach for it for multimodal perception
• Reach for it for agent planning
• Reach for it for tasks needing tool calling

50 hosts serve this model at the price above · the maker's documentation

The registry keeps the six most recent versions of a line. Older ones stay in the file and can be surfaced without a rebuild.

What is the context window size?

The context window is 1,000,000 tokens.

Does it support video input?

Yes, it accepts image, text, and video inputs, and returns text.

Is the model open weights?

Yes, the weights are open.