On-prem models / MiniMax
Run MiniMax-M3 on-prem
Native multimodal MoE (about 428B total, 23B active) trained on text, image and video from the start, with MiniMax Sparse Attention for million-token contexts. Supports enabled, adaptive and disabled reasoning modes.
Memory needed
| Precision | Weights | KV per request, 8K | KV per request, 32K | KV per request, 128K |
|---|---|---|---|---|
| BF16 (full precision) | 856 GB | 1.01 GB | 4.03 GB | 16.1 GB |
| FP8 | 428 GB | 0.50 GB | 2.01 GB | 8.05 GB |
| INT4 (AWQ / GPTQ) | 240 GB | 0.50 GB | 2.01 GB | 8.05 GB |
| GGUF Q4_K_M | 257 GB | 0.50 GB | 2.01 GB | 8.05 GB |
KV cache from the published config: 60 layers, 4 KV heads x 128 dims. 16-bit cache for BF16 rows, 8-bit otherwise.
Fit for ERP and manufacturing work
Technically capable for document-heavy ERP agents, but the licence adds paperwork for commercial deployment. Choose it only if its quality on your tasks clearly beats MIT/Apache alternatives.
Strengths
- Strong long-horizon coding and 'cowork' agent performance per the card
- Adaptive reasoning mode saves tokens on easy requests
- Image and video input
- Supported by vLLM, SGLang, KTransformers and AMD ATOM (MXFP4/MXFP8)
Watch out for
- Licence is non-commercial by default with notice and revenue conditions; review before production
- About 855 GB in BF16 (separate MXFP8 repo)
- GQA KV on all 60 layers grows with context despite sparse reads
GPUs that run it
One user, 8K context, 4-bit
| Hardware | Count | Memory used | Tok/s per user |
|---|---|---|---|
| NVIDIA B300 (Blackwell Ultra) 270GB | 1 | 242 / 243 GB | 194 |
| NVIDIA H200 SXM 141GB | 2 | 244 / 253 GB | 164 |
| NVIDIA H200 NVL 141GB | 2 | 244 / 253 GB | 164 |
| AMD Instinct MI355X 288GB | 1 | 242 / 259 GB | 194 |
| NVIDIA A100 40GB PCIe | 8 | 253 / 288 GB | 124 |
| NVIDIA A100 80GB PCIe | 4 | 247 / 288 GB | 120 |
| NVIDIA A100 80GB SXM | 4 | 247 / 288 GB | 123 |
| NVIDIA H100 PCIe 80GB | 4 | 247 / 288 GB | 122 |
20 people at once, 16K context, FP8
| Hardware | Count | Memory used | Tok/s per user |
|---|---|---|---|
| AMD Instinct MI210 64GB | 8 | 461 / 460 GB | 81 |
| AMD Instinct MI325X 256GB | 2 | 452 / 460 GB | 105 |
| Intel Gaudi 3 128GB (HL-325L OAM) | 4 | 455 / 460 GB | 102 |
| Apple Mac Studio M3 Ultra 512GB | 1 | 450 / 460 GB | 11 |
| Apple Mac Studio M5 Ultra 512GB | 1 | 450 / 460 GB | 16 |
| NVIDIA B300 (Blackwell Ultra) 270GB | 2 | 452 / 486 GB | 127 |
| NVIDIA H200 SXM 141GB | 4 | 455 / 507 GB | 118 |
| NVIDIA H200 NVL 141GB | 4 | 455 / 507 GB | 118 |
Serving and licence
SGLang cookbook, vLLM recipe, KTransformers tutorial and ROCm ATOM guide; MiniMaxAI/MiniMax-M3-MXFP8 roughly halves memory. Recommended temperature 1.0, top_p 0.95.
MiniMax Community License. Free for non-commercial use; commercial use requires a 'Built with MiniMax M3' notice and a one-time notice email, plus prior written authorisation above US$20M yearly revenue; military use is prohibited.
Design a MiniMax-M3 deployment
Preloaded with MiniMax-M3. Add your ERP and documents, change GPUs or users, and export the diagram.
Questions
How much VRAM does MiniMax-M3 need?
The weights alone take about 856 GB in BF16, 428 GB in FP8 and 240 GB in 4-bit. On top of that each request holds a KV cache of about 0.50 GB at 8K tokens of context (FP8 cache), so 20 people generating at once need roughly 10.1 GB more, plus a few GB of runtime overhead.
What is the smallest GPU setup that runs MiniMax-M3?
For a single user at 8K context in 4-bit, 1 x NVIDIA B300 (Blackwell Ultra) 270GB (270 GB each) fits, at an estimated 194 tokens per second. For a team, size for concurrent requests and context; the builder above does that.
Is MiniMax-M3 allowed for commercial use?
MiniMax-M3 is released under MiniMax Community License. Free for non-commercial use; commercial use requires a 'Built with MiniMax M3' notice and a one-time notice email, plus prior written authorisation above US$20M yearly revenue; military use is prohibited. Check the licence text on the model card with your legal team before production use.
Which engine should serve MiniMax-M3 on-prem?
SGLang cookbook, vLLM recipe, KTransformers tutorial and ROCm ATOM guide; MiniMaxAI/MiniMax-M3-MXFP8 roughly halves memory. Recommended temperature 1.0, top_p 0.95.