China AI Hub
Models / MiniMax-M3
MiniMax-M3
MiniMax M-series family · Provider: minimax · Status: active · Released: 2026-06-01
MiniMax-M3 is a MiniMax M-series model developed by minimax, a Chinese AI company headquartered in Room 1704-1, No. 1699 Gubei Road, Minhang District, Shanghai, China, released 2026-06-01 with a 1,048,576-token context window and open weights.
| Key fact | Value |
|---|---|
| Model ID | minimax-m3 |
| Architecture | Mixture-of-Experts: ~428B total / ~23B activated; MiniMax Sparse Attention (MSA) with claimed 9x prefill and 15x decode speedup vs M2 at 1M context |
| Context window | 1,048,576 tokens |
| Open weights | Yes |
| License | MiniMax Community License (custom): free for non-commercial use; commercial use requires prominent attribution 'Built with MiniMax M3' plus written authorization from MiniMax if yearly revenue exceeds US$20M (otherwise a one-time notice to api@minimax.io) |
| Self-hosting | Yes |
| API available | Yes |
| API pricing | $0.3 input / $1.2 output per 1M tokens (USD) · provider pricing page |
| Regions | china, international |
| Cloud providers | MiniMax Platform |
Capabilities
| Capability | Supported |
|---|---|
| Reasoning | Yes |
| Coding | Yes |
| Math | Unknown |
| Chinese | Unknown |
| English | Unknown |
| Multilingual | Unknown |
| Vision | Yes |
| Audio | Unknown |
| Video | Yes |
| Tool calling | Yes |
| Function calling | Unknown |
| Structured output | Unknown |
| Agent capability | Yes |
| RAG | Unknown |
| Computer use | Unknown |
Benchmark results
| Benchmark | Version | Score | Metric | Date | Source type | Source |
|---|---|---|---|---|---|---|
| BrowseComp | — | 83.5 | accuracy | 2026-06-01 | vendor_reported | link |
| PostTrainBench | — | 37.1 | accuracy | 2026-06-01 | vendor_reported | link |
| SWE-bench Pro | — | 59 | accuracy | 2026-06-01 | vendor_reported | link |
| Terminal-Bench 2.1 | — | 66 | accuracy | 2026-06-01 | vendor_reported | link |
| MCP Atlas | — | 74.2 | accuracy | 2026-06-01 | vendor_reported | link |
Benchmark scores are single data points, not universal rankings. Vendor-reported scores are labeled as such.
Known limitations
- Max output tokens not publicly disclosed in official docs
- 1M context with at least 512K guaranteed usable; pricing splits at the 512K input boundary
- Thinking is disabled by default (thinking=adaptive enables it)
- Benchmarks vendor-reported; not independently verified
MiniMax-M3 (2026-06-01) is MiniMax’s flagship model: a ~428B-total / ~23B-activated MoE with MiniMax Sparse Attention, a 1M-token context window (at least 512K guaranteed usable), and native multimodal input - text, image (JPEG/PNG/GIF/WEBP up to 10MB) and video (up to 50MB direct, 512MB via Files API) - with text output. It targets agentic reasoning, tool use, coding and long-context work.
Open weights (MXFP8, ~171k downloads) are on Hugging Face/GitHub under the MiniMax Community License. International pay-as-you-go pricing (standard tier, permanent 50% off vs list): $0.30 input / $1.20 output per 1M tokens up to 512K input (double above 512K), cache reads at $0.06; priority tier costs 1.5x (as of 2026-09-20). Max output tokens are not publicly disclosed.
Provider
Pricing
Benchmarks with results for this model
Agents built on this model
API
Data interpretation
| Field | Value | Evidence type |
|---|---|---|
| Context window | 1,048,576 tokens | Official |
| Architecture | Mixture-of-Experts: ~428B total / ~23B activated; MiniMax Sparse Attention (MSA) with claimed 9x prefill and 15x decode speedup vs M2 at 1M context | Official |
| Open weights | Yes | Official |
| License | MiniMax Community License (custom): free for non-commercial use; commercial use requires prominent attribution 'Built with MiniMax M3' plus written authorization from MiniMax if yearly revenue exceeds US$20M (otherwise a one-time notice to api@minimax.io) | Official |
| API pricing | $0.3 / $1.2 per 1M tokens (USD) | Official |
| BrowseComp | 83.5 accuracy | Vendor-reported |
| PostTrainBench | 37.1 accuracy | Vendor-reported |
| SWE-bench Pro | 59 accuracy | Vendor-reported |
| Terminal-Bench 2.1 | 66 accuracy | Vendor-reported |
| MCP Atlas | 74.2 accuracy | Vendor-reported |
| Family position | 1 model in the MiniMax M-series family | China AI Hub analysis |
Evidence types: Official = vendor documentation, pricing pages or model cards. Vendor-reported = benchmark scores published by the vendor. China AI Hub analysis = derived from the database itself. See the sourcing policy.
Sources
Confidence and source hierarchy per the sourcing policy. Facts change; check the source before relying on this page.
What is the context window of MiniMax-M3?
MiniMax-M3 has a 1,048,576-token context window.
Is MiniMax-M3 open weight?
Yes — MiniMax-M3 weights are openly available under the MiniMax Community License (custom): free for non-commercial use; commercial use requires prominent attribution 'Built with MiniMax M3' plus written authorization from MiniMax if yearly revenue exceeds US$20M (otherwise a one-time notice to api@minimax.io) license.
How much does MiniMax-M3 cost through the API?
$0.3 per 1M input tokens and $1.2 per 1M output tokens (USD).
Where does China AI Hub get its MiniMax-M3 data?
From 4 sources (official pages first), last verified 2026-09-20. Benchmark scores are labeled by source type; see the sourcing policy for details.