MiMo V2.5 Pro
MiMo V2.5 Pro needs roughly 699.7 GB VRAM at Q4_K_M quantization (2288.8 GB at FP16). 1 GPUs we track can run it fully in VRAM at 8k context.
1 GPU runs this natively · 0 with CPU offload
MiMo V2.5 Pro is a Mixture of Experts (MoE) large language model with 1020B total parameters but only 42B active per token developed by Xiaomi. Released in April 2026, it is a text-only model with a 1M context window, released under the MIT license, allowing commercial use. Trillion-parameter bilingual agent model with 384 routed experts, 8 selected per token, a 6:1 sliding/global-attention pattern, 128-token sliding windows, and three MTP layers. Xiaomi reports MMLU-Pro 68.5, GPQA-Diamond 66.7, SWE-bench Verified 78.9, and SWE-bench Pro 57.2. The official open checkpoint is FP8 and the full 1M context is extremely memory-intensive; generic low-bit rows are estimates for community conversions, not Xiaomi-provided GGUF files. The API launched April 23 and the weights were opened June 29.
To run MiMo V2.5 Pro locally, you need approximately 699.7 GB of VRAM at Q4_K_M quantization with 8k context. 1 of the GPUs we track can run it fully in VRAM, with a further 0 able to offload to system RAM. At Q4_K_M it requires 699.7 GB — more than any single consumer GPU. A multi-GPU server or 80 GB datacenter GPU is required. At Q8_K_M (1218.3 GB), you get near-FP16 quality while still fitting on large server or multi-GPU setups. FP16 requires 2288.7 GB, limiting it to datacenter-class hardware with 80 GB+ VRAM. As a MoE model, inference speed depends on active parameters (42B) rather than total size, so once it fits in VRAM it runs noticeably faster than dense models of the same parameter count.
With an MMLU-Pro score of 68.5, it delivers strong general reasoning for local deployment. The license allows commercial use.
VRAM at each quantization
MiMo V2.5 Pro natively supports a longer context window, but the table below is capped at 8k for comparability — KV cache grows linearly with context length.
| Quant | Weights | KV cache | Total |
|---|---|---|---|
| FP32 | 4080.0 GB | 3.52 GB | 4573.6 GB |
| BF16 | 2040.0 GB | 3.52 GB | 2288.8 GB |
| FP16 | 2040.0 GB | 3.52 GB | 2288.8 GB |
| Q8_0 | 1084.3 GB | 3.52 GB | 1218.3 GB |
| Q6_K | 837.4 GB | 3.52 GB | 941.9 GB |
| Q5_K_M | 726.2 GB | 3.52 GB | 817.3 GB |
| Q4_K_M | 621.2 GB | 3.52 GB | 699.7 GB |
| Q3_K_M | 490.6 GB | 3.52 GB | 553.4 GB |
| Q2_Krec | 388.6 GB | 3.52 GB | 439.2 GB |
| NVFP4cuda | 510.0 GB | 3.52 GB | 575.1 GB |
KV cache is calculated at 8k context (FP16). Note that NVFP4 only runs on CUDA GPUs. Turn on TurboQuant in the calculator above for lower KV cache estimates.
Benchmarks
GPUs that run MiMo V2.5 Pro natively (1)
- Apple M3 Ultra (512GB)Q2_K · 11.5 t/s
Notes
Trillion-parameter bilingual agent model with 384 routed experts, 8 selected per token, a 6:1 sliding/global-attention pattern, 128-token sliding windows, and three MTP layers. Xiaomi reports MMLU-Pro 68.5, GPQA-Diamond 66.7, SWE-bench Verified 78.9, and SWE-bench Pro 57.2. The official open checkpoint is FP8 and the full 1M context is extremely memory-intensive; generic low-bit rows are estimates for community conversions, not Xiaomi-provided GGUF files. The API launched April 23 and the weights were opened June 29.
How to run MiMo V2.5 Pro locally
Q2_K needs 439.2 GB — needs multiple datacenter-class GPUs (80 GB+ each).
vLLM (official FP8 checkpoint)
vllm serve XiaomiMiMo/MiMo-V2.5-Pro --trust-remote-codeLM Studio: The official release is FP8 and depends on custom model code. Community low-bit conversions may appear, but they are not equivalent to Xiaomi's supported checkpoint and need hundreds of gigabytes.
Why this quantization? Q2 is shown as the lowest remotely feasible generic local estimate for a 1.02T model, not as Xiaomi's preferred format. The supported checkpoint is FP8; serious deployment should use that checkpoint across a large GPU cluster and treat the calculator's GGUF rows as planning estimates only.
Who is MiMo V2.5 Pro for?
Large inference clusters serving bilingual long-context coding and agent workloads, not individual desktop users.
Best for
- Million-token repository and document agents
- Chinese-English coding and reasoning
- Long-running tool-driven workflows
Not ideal for
- Single-node consumer hardware
- Users seeking an officially supported GGUF
- Full-context use without substantial KV-cache headroom
Frequently asked questions
- What are the VRAM requirements for MiMo V2.5 Pro?
- MiMo V2.5 Pro requires approximately 699.7 GB of VRAM at Q4_K_M quantization, 1218.3 GB at Q8, and 2288.7 GB at FP16. These numbers assume 8k context window; VRAM scales linearly with context length due to the KV cache.
- How many parameters does MiMo V2.5 Pro have?
- MiMo V2.5 Pro has 1020 billion total parameters, but only 42 billion are active per token thanks to its Mixture of Experts (MoE) architecture. This makes inference significantly faster than the total parameter count suggests.
- How capable is MiMo V2.5 Pro?
- With an MMLU-Pro score of 68.5, MiMo V2.5 Pro delivers solid general-purpose performance suitable for most everyday tasks and professional use.
- Can MiMo V2.5 Pro run on a 16 GB GPU?
- No. At Q4_K_M, MiMo V2.5 Pro needs 699.7 GB of VRAM — more than 16 GB. You will need a multi-GPU server.
- Can MiMo V2.5 Pro run on a 24 GB GPU?
- No. Even at Q4_K_M, MiMo V2.5 Pro needs 699.7 GB. Consider a multi-GPU server with 80 GB+ total VRAM.
- What is the smallest quantization for MiMo V2.5 Pro that fits in 24 GB of VRAM?
- MiMo V2.5 Pro cannot fit in 24 GB of VRAM at any standard quantization level. The minimum needed is 439.2 GB at Q2_K.
- What GPU do I need to run MiMo V2.5 Pro locally?
- You need a multi-GPU server. At Q4_K_M, MiMo V2.5 Pro needs 699.7 GB VRAM, more than any single consumer GPU. Consider 2–4× H100 or A100 GPUs.