Runs on Multi-GPU · Hugging Face ↗
AA coding index 60.2 (Artificial Analysis, frozen since 2026-09-11)
Xiaomi's ~1-trillion-parameter MoE flagship — a strong open coder at the very top of the size range.
Real GGUF file sizes = weight VRAM. Add the KV cache for your context (≈2 GB at 32K, fp16). Read from the GGUF header, but this architecture has never been checked against llama.cpp — treat it as an estimate.
| Quantization | Size |
|---|---|
| IQ1_M | 304.1 GB |
| IQ2_XXS | 317.2 GB |
| IQ2_M | 317.3 GB |
| Q2_K_XL | 338.4 GB |
| IQ3_S | 377.9 GB |
| IQ3_XXS | 412.7 GB |
| Q3_K_M | 459.7 GB |
| Q3_K_XL | 459.9 GB |
| IQ4_XS | 490.7 GB |
| IQ4_NL | 501.3 GB |
| Q4_K_S | 588.5 GB |
| MXFP4 | 610.5 GB |
| Q4_K_M | 629.6 GB |
| Q4_K_XL | 631.0 GB |
| Q5_K_S | 713.2 GB |
| Q5_K_M | 758.0 GB |
| Q5_K_XL | 759.3 GB |
| Q6_K | 846.6 GB |
| Q6_K_XL | 928.5 GB |
| Q8_0 | 1087.6 GB |
| Q8_K_XL | 1101.5 GB |
| BF16 | 2046.7 GB |
~630 GB at Q4 — multi-node only. Serve with vLLM / SGLang tensor-parallel. Another "open but data-centre" model; rent it per token unless you have a fleet.