MiniMax M3
MiniMax-M3 is MiniMax’s latest M-series multimodal foundation model, built for agentic reasoning, tool use, coding, and long-context tasks. It supports text, image, and video inputs with text output, offering a 1M-token context window, extended thinking, function calling, and structured outputs. With strong capabilities in long-horizon agent workflows, software development, multimodal understanding, and extended response generation, MiniMax-M3 is ideal for autonomous agents, coding assistants, document and video analysis, and production-grade applications that require massive context at a competitive cost.
Related Models
MiniMax M2.7
MiniMax-M2.7 is a next-generation large language model built for autonomous real-world productivity and continuous improvement. Through advanced agentic capabilities and multi-agent collaboration, it can plan, execute, evaluate, and refine complex tasks in dynamic environments while actively contributing to its own evolution. Optimized for production-grade workflows, M2.7 excels at live debugging, root-cause analysis, financial modeling, and end-to-end document creation across Word, Excel, and PowerPoint. It achieves 56.2% on SWE-Pro, 57.0% on Terminal Bench 2, and a 1495 ELO rating on GDPval-AA, establishing a new benchmark for multi-agent systems operating in real-world digital workflows.
Read Me
MiniMax M3 API
MiniMax M3 is MiniMax's open-weight frontier model for coding, long-running Agents, and native multimodal understanding. Its Mixture-of-Experts architecture contains approximately 428B parameters while activating about 23B per token, combining a large model capacity with a smaller active compute footprint.
On iCreat, MiniMax-M3 is available through an OpenAI-compatible Chat Completions API. Applications can send text, image, and video inputs, stream responses, configure thinking, and connect function tools without operating the model's serving infrastructure.
Model Positioning
MiniMax M3 brings three capabilities into one model: frontier-oriented coding, a 1M-token context window, and native multimodality. It is designed for workloads that must keep source code, documents, visual evidence, tool results, and long interaction histories available within the same task.
M3 replaces the text-only M2.7 generation with a larger multimodal architecture and MiniMax Sparse Attention. It is especially relevant when an application needs whole-repository context, long video understanding, or extended Agent execution without moving every request to a premium proprietary model.
Key Capabilities
Long-Horizon Coding and Agents
M3 is trained for repository work, terminal execution, tool use, debugging, and iterative engineering. Interleaved Thinking lets it reassess the task between tool calls instead of reasoning only once before execution begins.
Native Image and Video Understanding
Text, image, and video were mixed from the beginning of training. The model can connect visual information with instructions, code, documents, and interaction state rather than treating vision as a separate late-stage capability.
Million-Token Context
The 1M-token context window can accommodate large repositories, extensive document sets, long videos, logs, and multi-step Agent histories. MiniMax recommends a 128K output budget for typical requests and allows up to 512K output tokens.
MiniMax Sparse Attention
MSA selects relevant key-value blocks instead of applying full attention across every token. MiniMax reports that this reduces long-context compute while preserving model quality, making million-token tasks more practical.
Configurable Thinking
M3 supports enabled, adaptive, and disabled thinking. Adaptive mode decides when additional reasoning is useful; disabled mode reduces latency for direct conversation and code-completion workloads.
Pricing
| Token Type | Price |
|---|---|
| Input | $0.60 per 1M tokens |
| Output | $2.40 per 1M tokens |
| Cache Read | $0.12 per 1M tokens |
Use Cases
- Whole-repository engineering: inspect architecture, trace cross-file behavior, implement features, and diagnose failures with more of the codebase in context.
- Multimodal software work: turn screenshots, interface recordings, diagrams, and written requirements into implementation or debugging tasks.
- Long-running Agent workflows: combine planning, function tools, intermediate evidence, and repeated verification across extended task histories.
- Long-video and document analysis: connect events across lengthy recordings, specifications, reports, and supporting materials.
- High-volume coding assistance: handle code explanation, completion, migration, test generation, and bounded fixes with configurable thinking.
Tool execution remains controlled by the application. Production systems should restrict permissions, validate outputs, and require approval before files, external services, or consequential environments are changed.
Model Comparison
MiniMax M3 vs MiniMax-M2.7
| Factor | MiniMax M3 | MiniMax-M2.7 |
|---|---|---|
| Positioning | Multimodal coding and Agent model | Text engineering and Agent model |
| Parameters | ~428B total; ~23B active | 229.9B total; ~9.8B active |
| Context Window | 1M tokens | 204,800 tokens |
| Maximum Output | 512K tokens | 200K tokens |
| Official Input Modalities | Text, image, and video | Text |
| Attention Architecture | MiniMax Sparse Attention | Grouped-Query Attention |
| Reasoning Control | Enabled, adaptive, disabled | Thinking always enabled |
| Core Difference | Native multimodality and 1M context | Smaller text-only model |
| Best Fit | Multimodal long-horizon Agents | Text engineering workflows |
MiniMax M3 vs Claude Fable 5 and GPT-5.6 Sol
| Factor | MiniMax M3 | Claude Fable 5 | GPT-5.6 Sol |
|---|---|---|---|
| Positioning | Open-weight multimodal coding model | Premium long-running Agent model | Premium tool-oriented reasoning model |
| Context Window | 1M tokens | 1M tokens | 1.05M tokens |
| Maximum Output | 512K tokens | 128K tokens | 128K tokens |
| Official Input Modalities | Text, image, and video | Text and image | Text and image |
| Reasoning Control | Enabled, adaptive, disabled | Always-on Adaptive Thinking | none to max; optional Pro mode |
| Weight Availability | MiniMax Community License | Proprietary | Proprietary |
| Core Difference | Native video and longest output | Always-on adaptive reasoning | Max effort, Pro mode, broad tool orchestration |
| Best Fit | Multimodal coding and long Agent sessions | Quality-first long Agents and knowledge work | Complex coding, research, and tool workflows |
Why Choose MiniMax M3?
Choose MiniMax M3 when one workflow must combine code, long context, and visual evidence. Its 1M context, 512K maximum output, native video input, and sparse-attention architecture give developers room for tasks that outgrow conventional coding assistants.
iCreat provides a managed API route to these capabilities. Teams can validate prompts in the Playground, then use the same model ID with familiar OpenAI SDK patterns for streaming, thinking, and tool-connected applications.
Specifications
| Category | Description |
|---|---|
| Model Name | MiniMax M3 |
| Provider | MiniMax |
| Model ID | MiniMax-M3 |
| Release Date | June 1, 2026 |
| Model Type | Native multimodal MoE reasoning model |
| Context Window | 1M tokens |
| Maximum Output | 512K tokens |
| Knowledge Cutoff | Not published by MiniMax |
| Official Input Modalities | Text, image, and video |
| Output Modalities | Text |
| Official Reasoning Control | Enabled, adaptive, disabled |
| Default Effort | Adaptive thinking |
| Total Parameters | Approximately 428B |
| Activated Parameters | Approximately 23B per token |
| Attention Architecture | MiniMax Sparse Attention |
| Weight Availability | MiniMax Community License |
| iCreat Input Modalities | Text, image, and video |
| iCreat API Features | OpenAI-compatible Chat Completions, streaming, thinking, tool workflows |
| Best Suited For | Multimodal coding, long-context analysis, long-running Agents |
Architecture
MiniMax M3 is a 60-layer multimodal Mixture-of-Experts model with approximately 428B total parameters and 23B activated per token. Its text backbone uses 128 routed experts with top-four routing, while a vision encoder supplies image and video representations to the language model.
MiniMax Sparse Attention adds a lightweight index branch that selects relevant key-value blocks for each attention group. The main branch then applies exact attention only to those blocks. This content-dependent selection reduces the compute and memory cost of million-token inference without relying on a fixed local window.
Production Notes
Treat the 1M context window as a shared request envelope. Reserve space for system instructions, multimodal inputs, tool definitions, thinking, tool results, and final output; do not assume the entire window is available for source material.
For ordinary requests, start below the recommended 128K output budget and raise the limit only when the deliverable requires it. Use streaming for long responses so users can see progress and applications can detect early failures.
When a tool workflow uses Interleaved Thinking, retain the complete assistant message, including reasoning details, before appending tool results. Removing or rewriting this state can break reasoning continuity across tool rounds.
Validate tool-result bindings, especially when the same tool is called in parallel with different arguments. Echo identifying fields in tool responses and verify the final mapping before taking action.
FAQ
Does reasoning_effort: "low" turn thinking off?
No. In the iCreat example, reasoning_effort: "low" is used together with thinking: {"type": "enabled"}. The thinking mode controls whether M3 reasons, while effort configures the reasoning budget exposed by the endpoint.
Do I need to configure prompt caching to receive the Cache Read rate?
No separate cache flag is required. M3 supports automatic prompt caching; the Cache Read rate applies when a request reuses an eligible cached prefix. Keep stable instructions and reusable context at the beginning of repeated requests to improve the chance of a cache hit.
Does video input mean MiniMax M3 can generate videos?
No. Video is an input modality for understanding and analysis; MiniMax M3 returns text. Video generation requires a separate video-generation model or service.
Does the MiniMax Community License affect calls made through iCreat?
The Community License governs independent use and deployment of the published model weights. Calling MiniMax M3 through iCreat is a managed API workflow and does not require downloading or self-hosting those weights; API usage remains subject to iCreat's applicable terms.