| Model profile | ||
|---|---|---|
| Model | Size | Approximate size |
| MiMo V2.6 FlashModel source: provider1,048,576 tokens max native context | 309Bactive 15B | 185.4 GB4-bit quantization |
| Compatible hardware Estimated speeds, not benchmark results: calculated from memory bandwidth and model size. Real results can differ significantly because there is no precise formula for deriving LLM generation speed from hardware specifications alone. | ||||
|---|---|---|---|---|
| Hardware | Memory | Bandwidth | vLLM | Estimated generation |
| M5 Ultra | 256 GBUnified RAM | 1200 GB/s | no | 51 tok/s4-bit quantization |
| DGX H200 | 1128 GBGPU HBM3e | 4800 GB/s | yes | 167 tok/s4-bit quantization |
| DGX Station | 748 GBCoherent Memory | 7100 GB/s | yes | 221 tok/s4-bit quantization |
| ET900N G3 | 748 GBCoherent Memory | 7100 GB/s | yes | 221 tok/s4-bit quantization |
| DGX B200 | 1440 GBGPU HBM3e | 8000 GB/s | yes | 239 tok/s4-bit quantization |
| GB200 NVL72 | 13400 GBGPU HBM3e | 8000 GB/s | yes | 239 tok/s4-bit quantization |
| GB300 NVL72 | 20000 GBGPU HBM3e | 8000 GB/s | yes | 239 tok/s4-bit quantization |
Teardown of Xiaomi MiMo V2.6 Pro and Flash configs reveals MoE routing, native 4-bit training, MTP layers, DFlash, GQA and KV cache mechanics enabling 1M context within 527 GB.
Tests Xiaomi Mimo V2.6 Pro and Flash on website generation, C++ development, 3D modeling, Chrono City generation, RuneSClone, FPS development, and skate games.
Tests evaluate MiMo-V2.6-Flash on coding, secure code review, prompt-injection resistance, and prompt refusals against its benchmark performance.
MiMo V2.6 Flash RL Q9 runs locally on Mac Studio M3 Ultra, covering coding, safety tests, multimodal prototypes, simulations, games, web apps, logic, math and final verdicts.