Use this page as the stable entry point for Xiaomi MiMo V2.5 Pro work on RTX PRO 6000 Blackwell.
| Need | Page |
|---|---|
| Run MiMo V2.5 Pro FP4-DFlash on vLLM | Xiaomi MiMo V2.5 Pro FP4-DFlash v3 |
| Reproduce the previous vLLM DFlash page | MiMo FP4-DFlash v2 |
| Understand the older SGLang/B12X route | MiMo V2.5 Pro hub |
| Reproduce quantization details | MiMo quantization |
| Reproduce SGLang launch details | MiMo SGLang running notes |
| Area | Current guidance |
|---|---|
| Current model | XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash |
| Current runtime | vLLM V2 with DFlash |
| Attention | TRITON_ATTN for the documented v3 path |
| MoE | FlashInfer CUTLASS path in the documented launch |
| Long context | Do not set --max-model-len unless the page explicitly says to; the checkpoint advertises its own long context length. |
| Page | Status | Why keep it |
|---|---|---|
| MiMo FP4-DFlash v3 | Current | Fathomless validation and seq-mask fix notes. |
| MiMo FP4-DFlash v2 | Historical | Earlier vLLM DFlash runbook. |
| MiMo FP4-DFlash | Historical | Initial MiMo DFlash page. |
| MiMo V2.5 Pro SGLang/B12X hub | Archive | Older SGLang and B12X integration path. |
- Check for the fast DFlash marker in logs. A
diffkvattention marker usually means the wrong path is selected for the current MiMo fast recipe. - MiMo has historically been sensitive to attention-mask argument ordering, so smoke-test short decode and acceptance before running a long benchmark.