Xiaomi's MiMo-V2.5 is a single 310B MoE release at 1M context, MIT licensed. A datacenter-class model; local use means heavy offload or multi-GPU rigs.
MiMo is Xiaomi's open-weight model family, and right now it is a family of one: MiMo-V2.5, released in 2026 under the MIT license. It is a 310B-parameter Mixture-of-Experts model with 256 experts, 8 routed per token, and only 15B parameters active per forward pass. That combination — datacenter-scale total weights, phone-call-scale active compute — is exactly what MoE is for, and it makes MiMo-V2.5 one of the most interesting "big brain, modest appetite" models you can self-host. The catch: those 310B parameters still have to live somewhere, and that somewhere is your VRAM.
This page walks through what MiMo-V2.5 actually demands from your hardware, using the same VRAM formula as our GPU guides: Q4 at 8K context ≈ Q4 weight size + 0.025 × active params (B) × 8 + 1 GB overhead.
Size ladder
- MiMo-V2.5 (310B total, MoE with 15B active) — Q4 weights alone are ~171 GB; at 8K context the full working set is roughly 174 GB → you want a 192 GB+ rig (think 2×96 GB or 8×24 GB) to run it comfortably. FP16 weights are about 620 GB, firmly multi-node territory. See the full estimate breakdown on the MiMo-V2.5 model page.
There is no smaller sibling yet — if 192 GB of VRAM is outside your budget, MiMo is a family to admire from the cloud rather than run at home. The Q2 quant (~87 GB of weights) is the one partial escape hatch: it squeezes onto a 96 GB card or a dual-48 GB setup, though at that bitrate you should expect a visible quality drop on a model this size.
What makes the family special
- Extreme sparsity. Only 15B of 310B parameters are active per token (~4.8%), so per-token speed is closer to a 15B dense model than to anything in the 300B class. The bottleneck is memory capacity and bandwidth, not compute.
- 1M context window. The model supports up to 1024K context. Be warned that KV cache scales with context length: our formula's 8K baseline adds only ~3 GB, but pushing anywhere near the full 1M window adds hundreds of GB on top of the weights. Long-context MiMo is a server-grade workload.
- MIT license. Fully permissive — commercial use, fine-tuning, and redistribution are all on the table, which is rare at this parameter count.
- Single flagship strategy. Xiaomi shipped one very large model instead of a ladder of sizes, so there is no "MiMo-7B for your gaming GPU" option. Plan your hardware around the flagship or look at other families.
Which one should you pick
There is only one model, so the real question is which GPU tier makes it viable:
- 12 GB / 16 GB / 24 GB cards: not realistic for MiMo-V2.5, even with aggressive offload — you would be paging hundreds of gigabytes from system RAM or SSD and measuring output in seconds per token, not tok/s. Browse our best GPU picks for what actually works.
- 48 GB / 96 GB: still short of Q4. The Q2 quant (~87 GB weights) can technically fit on 96 GB with a small context window, and it is a fun experiment, but treat it as a preview of the model rather than a daily driver.
- 192 GB and up: this is the entry ticket for Q4 at a usable 8K context — dual 96 GB cards, a 4×48 GB workstation, or an 8×24 GB rig. Expect MoE-friendly throughput since only 15B parameters do the work per token.
- Multi-node / datacenter: required for FP16 (~620 GB) or for exploiting the 1M context window at scale.
The short version: MiMo-V2.5 is a phenomenal deal in capability-per-watt terms, but the cover charge is 192 GB of VRAM. If that is you, the MIT license and MoE speed make it one of the best self-hostable flagships of 2026.
Links
1 models in the MiMo family, grouped by series. Q4 (GB) is weights-only; total VRAM adds KV cache and overhead — the fit check links compute it for a representative retail GPU at 8K context.