MiniMax
MiniMax-M3
Released 
June 2026

MiniMax
MiniMax-M3

Open-weight model uniting frontier coding, 1M-token context, and native image/video input for long-horizon agents.

Modality:
Text
Image
PDF
model ID
minimax/minimax-m3

Output Speed *

120.69
tok/s

Intelligence Index *

29.6
/ 100

Context Window *

524288
tokens

Input price

1.8
Anytoken

Output price

7.2
Anytoken
MiniMax M3: Open-Weight Coding and Agentic Frontier at 1M-Token Context MiniMax M3 is an open-weight, natively multimodal MoE model from Shanghai-based MiniMax, sitting at the top of the current M-series above M2.7. It combines frontier-level coding, a 1M-token context window built on MiniMax Sparse Attention (MSA), and native image and video input in a single checkpoint. With roughly 428B total and ~23B activated parameters, it targets long-horizon agentic work: multi-file coding, tool-driven browsing, and desktop automation. The strongest fit is sustained, multi-step development and agent workflows where large context and low long-context compute cost matter more than single-turn latency. Integrate MiniMax M3 via the AnyAPI.ai unified API.

Performance

Where MiniMax M3 Earns Its Place: Long-Horizon Coding and Agents

MiniMax M3 is built for sustained, multi-step engineering and agent runs rather than single-turn chat. On MiniMax's own evaluations it posts 59.0% on SWE-Bench Pro and 74.2% on MCP Atlas, with strong tool-use and browsing results. Its defining trait is the MSA architecture, which cuts per-token compute at 1M tokens to roughly 1/20 of the previous generation with substantially faster prefill and decode. For production, this makes very long codebases, transcripts, and multi-document agent state economically viable to process, and enables autonomous runs measured in hours rather than single calls.

Benchmarks

MiniMax M3 Benchmarks: Coding, Tool Use, and Autonomous Browsing

Independent testing by Artificial Analysis places MiniMax M3 at 45 on its Intelligence Index, well above the ~28 median for comparable models, at around 114 tokens per second with a 1.34s time to first token. MiniMax's own benchmarks report 59.0% on SWE-Bench Pro, 66.0% on Terminal-Bench 2.1, and 83.5 on BrowseComp for autonomous browsing. Note that most coding and agent headline numbers are vendor-published and pending broad independent verification; premium closed models such as Claude Opus 4.8 still lead pure code-modification rows. Treat published scores as selection signals and validate on your own tasks.

Output Speed

*
120.69
tok/s

Intelligence Index

*
29.6
/ 100

MMLU *

Broad world knowledge and problem-solving
0
%

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.
93
%

HLE *

Adherence to multi-step structured instructions.
39
%

LiveCodeBench *

Tool-calling reliability in long agentic loops.
0
%

Technical Specifications

What the model supports

MiniMax M3 accepts text, image, and video input and returns text only — it is not an image, audio, or video generator. The API supports up to a 1M-token context window with a guaranteed 512K minimum, and a maximum output distinct from that context budget. A toggleable thinking mode lets you enable extended reasoning for complex tasks or disable it for latency-sensitive calls at the same pricing. The 1M context, powered by MSA, is the key production enabler: it makes whole-codebase and long-transcript workflows viable without prohibitive long-context compute.
Verified Specifications — 
MiniMax-M3
*
Input modalities
Text
Image
PDF
output modalities
Text
Context window
524288
 tokens
Maximum output tokens
512000
Reasoning
Yes
Knowledge cutoff
June 2026
Pricing (standard)
1.8
 AnyTokens in
 / 
7.2
 AnyTokens out

Comparison

MiniMax M3 vs MiniMax M2.7: What Changes for Long-Context Agents?

MiniMax M2.7 is M3's direct predecessor and remains the practical alternative for teams already integrated with the M-series. Both target coding and agent workloads, but the architecture differs sharply. M2.7 ships a roughly 200K-token context window and is text-focused with a highspeed variant; M3 introduces MSA, extending context to 1M tokens, adds native image and video input, and reports higher engineering-benchmark scores. On SWE-Bench Pro, M3's 59.0% edges past M2.7's 56.2%. The core decision is whether you need M3's much larger context and multimodality, or whether M2.7's smaller, stable, well-integrated context is sufficient.

Dimension
MiniMax-M3
MiniMax-M2.7
Context window *
524288
tokens
196608
tokens
Output speed *
120.69
tok/s
0.00
tok/s
Intelligence Index *
29.6
23.2
Input pricing
1.8
AnyToken
1.8
AnyToken
Output pricing
7.2
AnyToken
7.2
AnyToken
Knowledge cutoff *
June 2026
March 2026

Choose MiniMax M3 when you need whole-codebase or long-transcript context, native image/video understanding, or long-horizon autonomous agent runs where MSA's low long-context compute cost matters. Choose MiniMax M2.7 when your workflows fit comfortably within ~200K tokens, you are already integrated with its endpoints and highspeed variant, or you want the more established, text-only option without migrating SDK, billing, or endpoint configuration. M3 does not dominate every row — frontier closed models still lead some coding tasks — so validate against your own workloads before switching.

Limitations & Trade-offs

Where MiniMax-M3 falls short

1
Vendor-reported benchmarks. Most headline coding and agent scores — 59.0% SWE-Bench Pro, 66.0% Terminal-Bench 2.1, 83.5 BrowseComp — come from MiniMax's own evaluations and remain pending broad independent verification. Independent intelligence measurements are favorable, but for coding-agent selection this matters: benchmark ranking and real-world performance on your specific codebase are different questions. Teams building production coding or agent systems should run M3 on their own tasks before committing, especially where a premium closed model like Claude Opus 4.8 may lead pure code-modification work.
2
Text-only output and no audio input. MiniMax M3 accepts text, image, and video input but returns only text. It does not generate images, audio, or video, and audio input is not part of its verified modality set. Teams needing speech input or any generative media output must pair M3 with separate models (MiniMax's own H3 for video or Speech models, or third-party services). For voice agents or media-generation pipelines, M3 handles only the reasoning and text layer.
3
Long-context guarantees vary. While the API advertises up to 1M tokens, the guaranteed minimum is 512K, and effective ceilings can differ by host and tier. Third-party endpoints may also restrict tool calling — some OpenRouter routes, for example, do not accept tools. Teams relying on full 1M-token context or function calling should confirm request format, output-token limits, and tool support in the exact provider account and route they plan to use, rather than assuming uniform behavior across every host.
4
Licensing and self-hosting cost. M3's downloadable weights use a commercially restricted custom license rather than a standard permissive open-source license, so verify terms before building on the open weights. With ~428B total parameters, full-scale self-hosting also requires substantial infrastructure despite the ~23B activated-parameter MoE design. Teams wanting true open-source licensing or lightweight local deployment should confirm the license and hardware requirements before treating M3 as a drop-in open model.

Best-Fit Workloads

Where this model earns its place

01

Long-horizon coding agents


M3 is oriented toward multi-step, tool-driven development rather than single completions, with autonomous demonstrations running 12–24 hours — including a CUDA kernel optimization that improved hardware utilization across nearly 2,000 tool calls without human intervention. Combined with 1M-token context and reported 59.0% SWE-Bench Pro, this fits repository-wide refactoring, bugfixing, and implementation agents. Validate on your codebase, since vendor benchmarks are pending independent confirmation.

02

Whole-codebase and long-document RAG


The MSA architecture makes the 1M-token context economically usable, cutting per-token compute at long context to roughly 1/20 of the previous generation with much faster prefill and decode. This supports feeding entire codebases, large log sets, or multi-document retrieval packs in a single request without the prohibitive long-context cost of full attention. Ideal for code-understanding assistants and research pipelines that reason across large corpora.

03

Autonomous browsing and tool-use agents


M3 reports strong agentic tool-use results, including 74.2% on MCP Atlas and 83.5 on BrowseComp for autonomous web research, and it can operate a desktop computer. This suits research agents, browser automation, and office-workflow assistants that plan, invoke tools, and iterate. Confirm tool-calling support on your chosen endpoint, as some third-party routes restrict function calling.

04

Grounded multimodal analysis


Because M3 is natively multimodal from training step zero, it handles screenshots, documents, charts, and video frames alongside text in one workflow — independent profiling rates its multimodal and grounded category among its stronger areas. This fits UI analysis, chart-to-code, document review, and visual QA tasks. Remember output is text only, so pair M3 with a generative media model if you need image, audio, or video output.

Pricing in anytokens via AnyAPI
Input
1.8
Output
7.2
Cache write
Cache read

Integration

Access MiniMax-M3 via AnyAPI.ai

Access MiniMax-M3 through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate MiniMax-M3 without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access MiniMax-M3 and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test MiniMax-M3 against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use MiniMax-M3 from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use MiniMax-M3 for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

MiniMax M3 supports a context window of up to 1,048,576 tokens (1M), with a guaranteed minimum of 512K tokens, enabled by the MiniMax Sparse Attention (MSA) architecture. Its maximum output is reported at 262,144 tokens, which is separate from the input context budget. Effective limits can vary by host and tier, so confirm in your chosen provider account.

MiniMax M3 is positioned for frontier coding and agentic work. On MiniMax's own benchmarks it scores 59.0% on SWE-Bench Pro and 66.0% on Terminal-Bench 2.1, and it has run autonomous coding tasks for 12–24 hours. These are largely vendor-reported numbers pending independent verification, so validate on your own codebase before production use.

The MiniMax M3 API accepts text, image, and video input and returns text output only. It is natively multimodal, trained on interleaved data from the first step. It does not generate images, audio, or video, and audio input is not part of its verified modality set — pair it with dedicated media models for those needs.

Yes. MiniMax M3 supports a toggleable thinking mode that can be turned on for complex reasoning or off for faster, latency-sensitive responses, switched at request time without a separate model version. Both modes share the same pricing, so you can trade depth for speed per request depending on the task.

MiniMax M3 is open-weight, with weights released on Hugging Face and GitHub for self-hosting. However, it uses a commercially restricted custom license rather than a standard permissive open-source license, and with roughly 428B total parameters, full-scale self-hosting requires substantial infrastructure. Confirm the license terms before building on the open weights.

* Benchmark data source: Artificial Analysis artificialanalysis.ai