One-line summary
MiniMax officially open-sourced the M3.1 Flash model weights on October 23, 2026: 428B total parameters with 23B active parameters (MoE architecture); real users processed over 100 trillion tokens within 72 hours of launch; MiniMax states inference speed is nearly 3x faster and usage cost is 70% lower than the previous generation, with adaptation work across more than six chip vendors.
Data description
- Data sources: reports on MiniMax's official open-source announcement by Cailianshe (第一财经), National Business Daily (每日经济新闻) and Bianews (鞭牛士), all dated October 23, 2026; MiniMax Open Platform official documentation.
- Scope: the token figure covers the 72 hours after M3.1 Flash went live (launched on the evening of October 19, 2026); speed and cost figures are MiniMax's official October 19 release statements versus the previous-generation M3.
- As of: October 23, 2026.
Core data
Model overview
| field | Meaning | Value / note |
|---|---|---|
| model_name | Model name | MiniMax M3.1 Flash |
| total_parameters | Total parameters | 428B |
| active_parameters | Active parameters per token | 23B |
| architecture | Architecture | Mixture of Experts (MoE) with full-model sparse attention |
| announce_date | Launch date | October 19, 2026 |
| open_source_date | Weight open-source date | October 23, 2026 |
| license | License | Per official release (previous-gen M3 uses MiniMax Community License) |
| weight_repo | Weight repository | Hugging Face MiniMaxAI org (live after official release) |
Performance and cost
| field | Meaning | Value / note |
|---|---|---|
| tokens_processed_72h | Tokens processed in 72 hours after launch | Over 100 trillion |
| speed_improvement | Inference speed vs M3 | Nearly 3x (official figure) |
| cost_reduction | Usage cost vs M3 | 70% (official figure) |
| optimization_1 | Optimization 1 | Full-model sparse attention |
| optimization_2 | Optimization 2 | Quantization-aware training |
| optimization_3 | Optimization 3 | Power-aware scheduling |
| api_price_input | Platform API input price at launch | CNY 1 per million tokens |
| api_price_output | Platform API output price at launch | CNY 4 per million tokens |
Ecosystem adaptation
| field | Meaning | Value / note |
|---|---|---|
| chip_partner_count | Number of chip partners | More than 6 |
| chip_partners | Chip partners | Huawei Ascend, T-Head, Kunlunxin, Moore Threads, NVIDIA, AMD |
Field notes
- total_parameters vs active_parameters: in a MoE (Mixture of Experts) architecture, total parameters include all expert weights, while active parameters are what actually run for each token; the larger the gap, the lower the inference cost per token.
- tokens_processed_72h: total tokens real users processed on the open platform within 72 hours after M3.1 Flash went live on the evening of October 19, 2026.
- license: as of October 23, 2026, MiniMax had not separately announced the M3.1 Flash license in its open-source announcement; previous-gen MiniMax M3 BF16 and MXFP8 weights use the MiniMax Community License. Treat the actual license in the M3.1 Flash repository as authoritative.
- api_price: platform price at M3.1 Flash launch (per Tencent Cloud Developer Community, October 19, 2026); subsequent pricing per the open platform.
Data sources and references
- Cailianshe (第一财经): MiniMax M3.1 Flash processed over 100T tokens in three days; officially open-sourced today (Oct 23, 2026) link
- National Business Daily: MiniMax open-sources M3.1 Flash (Oct 23, 2026) link
- Bianews: MiniMax M3.1 Flash officially open-sourced (Oct 23, 2026) link
- Tencent Cloud Developer Community: MiniMax M3.1 Flash launch - CNY 1 in / CNY 4 out per million tokens (Oct 19, 2026) link
- MiniMax Open Platform Docs: Self-hosting MiniMax-M3 (M3 weights and license reference) link