Models › MiniMax M3
MiniMax · Model details, evidence, and API facts
MiniMax M3 is MiniMax’s open-weight flagship, the first model to combine frontier coding
Starting provider API rate; context tiers apply
MiniMax M3 is MiniMax’s flagship open-weight model, released June 1, 2026 as the first open-weight model to combine frontier-level coding performance
It is a highly impeccable model for high-quality performance
Only results mapped to this exact model and backed by a named source appear here. Different evaluation protocols are not treated as interchangeable.
| Benchmark | Score | Evaluation details | Source |
|---|---|---|---|
Artificial Analysis Intelligence Index A composite benchmark evaluating models across reasoning, knowledge, mathematics, and coding. |
29 Composite index score |
This composite benchmark evaluates models across reasoning, knowledge, mathematics, and coding · Harness: Artificial Analysis standardized harness | Methodology Artificial Analysis Independently reproduced |
SWE-Bench Pro A benchmark measuring real-world software-engineering fixes on production-scale repositories. |
59.0% Resolve rate |
Version: SWE-Bench Pro · surpasses GPT-5.5 and Gemini 3.1 Pro, and approaches Opus 4.7 · Reasoning: thinking enabled | Methodology MiniMax · Vendor-reported |
Terminal-Bench 2.1 Evaluates autonomous terminal use and multi-step command execution. |
66.0% Task success rate |
Version: Terminal-Bench 2.1 · Coding & Agentic Frontier launch result · Reasoning: thinking enabled | Methodology MiniMax · Vendor-reported |
MCP Atlas An end-to-end evaluation of tool-use and MCP-based agentic workflows. |
74.2% Task success rate |
Version: MCP Atlas · Coding & Agentic Frontier launch result · Reasoning: thinking enabled | Methodology MiniMax · Vendor-reported |
BrowseComp Tests autonomous web browsing and multi-step information retrieval. |
83.5 Score |
Surpasses Opus 4.7 (79.3), demonstrating strong autonomous browsing and information retrieval capabilities · Reasoning: thinking enabled | Methodology MiniMax · Vendor-reported |
Arena scores come from blind human preference votes. They are reported separately from task benchmarks and are not used as substitutes for missing benchmark results.
No exact Arena match is available for this model. Similar names and provider variants are not merged automatically.
Technical limits and reference provider prices come from the current catalog unless a reviewed external source is listed. Missing values stay marked as unavailable.
| Model | |
| Provider | MiniMax (Xiyu Technology) |
| Release date | June 1, 2026 |
| Catalog added | Sep 25, 2026 |
| Context window | 1,048,576 tokens |
| Maximum output | 262,144 tokens |
| Knowledge cutoff | Not available in reviewed sources |
| License | MiniMax Community License — commercial use allowed with restrictions; open weights |
| Capabilities and API | |
| Input types | File, Image, Video, Text |
| Output types | Text |
| Reasoning | Supported — configurable thinking (enabled, disabled, or adaptive) |
| Provider endpoint tool parameters | Accepts tool parameters; supports computer-use |
| Base provider API input | $0.30 / 1M tokens |
| Base provider cached input | $0.06 / 1M tokens |
| Base provider API output | $1.20 / 1M tokens |
| Provider tier at higher volume | Not published — no confirmed volume-based tiering in reviewed sources |
| Rates shown are MiniMax's direct API pricing. Third-party routing providers publish different rates for the same open-weight model — OpenRouter lists $0.23 / $0.96 per 1M input/output tokens, and DeepInfra lists $0.28 / $1.10 — so actual cost depends on which endpoint serves the request. These are reference provider API rates, not Pyxa AI checkout charges — access in Pyxa AI follows the allowances of your Pyxa AI plan. | |
Source links are attached to the facts and evaluations they support. Catalog-only values are not presented as independently verified claims.
Open another model’s detail page or compare it directly with GPT-5.6 Sol. Suggestions come from the current catalog.
Use this model, then compare its response with other AI models in the same workspace.