MiMo V2.5 Pro
XiaomiMiMo · released 2026-04-22 · mit license
MiMo-V2.5-Pro is Xiaomi's flagship open-source (MIT) MoE (1.02T total / 42B active, April 2026) for agentic and software-engineering tasks, scoring 54 on the Artificial Analysis Intelligence Index (#2 among open weights in its class).
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 1023.24B total · MoE, — active |
| Architecture | mimo_v2 |
| Context window | 1049K tokens |
| Knowledge cutoff | — |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | — |
|---|---|
| SWE-bench Verified | 78.9% |
| AIME | — |
| MMLU-Pro | — |
| BFCL v3 (tool use) | — |
| Composite score | 6.73 |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 595 GB | 593.5 GB | — | 1049K |
Strengths & weaknesses
Strengths: Among the leading open-weight models in intelligence (AA Index 54); Strong agentic/SWE performance (SWE-bench Verified 78.9); 1M context, MIT licensed
Weaknesses: Very large (1.02T total / 42B active), expensive to serve; Text-only (no multimodal), verbose, high TTFT