$ cat wiki/models/minimax-m3.md
MiniMax M3
modelupdated 2026-07-21created 2026-07-17
Spec
| Attribute | Value |
|---|---|
| Developer | [MiniMax](/wiki/entities/minimax) |
| Released | June 1, 2026 |
| Announced | June 1, 2026 (MiniMax official blog) |
| Architecture | MSA (MiniMax Sparse Attention) — sparse MoE |
| Total parameters | ~428B |
| Active parameters | ~23B per token |
| Context window | 1,000,000 tokens (1M); min. 512K guaranteed |
| Max output | 131,072 tokens |
| Modalities | Text, image, video (native multimodal, no post-hoc adapters) |
| License | Open-weight (HuggingFace) |
| API access | Global (including English) |
| Pricing | Available via API; pricing on OpenRouter |
| Availability | API (global, including English) + open-weight on HuggingFace |
Release Date
June 1, 2026 — MiniMax official blog and HuggingFace release. Captured by this wiki July 17 due to WAIC 2026 coverage surfacing the model.
Benchmarks
| Benchmark | MiniMax M3 | Comparison | Notes |
|---|---|---|---|
| SWE-Bench Pro | 59.0% | Beats GPT-5.5 (prior OpenAI frontier) | Coding agent |
| BrowseComp | 83.5 | Beats Claude Opus 4.7 (79.3) | Browser agent |
| OSWorld-Verified | 70.06% | — | Computer use |
| Note on comparisons: GPT-5.6 Sol (July 9, 2026) is the current OpenAI frontier; M3's comparisons are vs. GPT-5.5. Claude Sonnet 5 (June 30, 2026) surpasses Opus 4.7 on most benchmarks. M3's BrowseComp result is nonetheless impressive for a Chinese open-weight model. |
Architecture
- MSA (MiniMax Sparse Attention): proprietary sparse attention mechanism
-
9× prefill speedup at 1M-token context vs. M2 (prior MiniMax model)
-
15× decoding speedup at 1M context vs. M2
- 1/20th per-token compute vs. M2 at 1M context
-
- MoE: ~428B total, ~23B active — efficient for inference relative to total parameter count
Use Cases
- Agentic coding (SWE-Bench Pro 59.0%)
- Browser-based agent tasks (BrowseComp 83.5)
- Computer use (OSWorld 70.06%)
- Long-context document analysis (1M tokens)
- Multimodal workflows (text + image + video)
Significance
MiniMax M3 is notable as:
- One of the first Chinese open-weight models to surpass a leading US closed model (Claude Opus 4.7) on an agent benchmark (BrowseComp 83.5 vs. 79.3)
- A frontier-class open-weight model with 1M context and native multimodality from a non-Tier-1 (non-Alibaba/ByteDance) Chinese lab
- Available globally via API and open-weight — expanding Chinese AI model access internationally
Compared To
| Model | Org | SWE-Bench Pro | Notes |
|---|---|---|---|
| MiniMax M3 | MiniMax | 59.0% | Open-weight, 428B MoE, BrowseComp 83.5 |
| GPT-5.5 | OpenAI | ~58.6% | M3's reference at launch; now superseded by GPT-5.6 |
| Claude Opus 4.7 | Anthropic | ~64.3% | BrowseComp 79.3 (M3 beats it here) |
| GLM-5.2 | Z.ai | — | Comparable Chinese open-weight frontier |
| Claude Sonnet 5 | Anthropic | 63.2% | Current mid-tier Anthropic (GA June 30) |
Sources
- MiniMax official blog: https://www.minimax.io/blog/minimax-m3
- HuggingFace model: https://huggingface.co/MiniMaxAI/MiniMax-M3
- OpenRouter: https://openrouter.ai/minimax/minimax-m3
- Pandaily: https://pandaily.com/minimax-m3-model-2026
- MarkTechPost: https://www.marktechpost.com/2026/06/01/minimax-releases-minimax-m3-with-msa-architecture-supporting-1m-token-context-native-multimodality-and-agentic-coding/
Related
- MiniMax — developer
- Z.ai — comparable Chinese frontier lab (GLM-5.2, 744B MoE)
- Agents (LLM Agents) — agent benchmark performance anchor
- Reasoning Models — 1M context + long-horizon tasks