Can I run Inkling-Small on an NVIDIA B200?
Yes, with a GGUF version
The published weights are too big, but the MXFP4_MOE file from unsloth/Inkling-Small-GGUF (158 GB) fits, with room for ≈ 132K tokens of context.
Full analysis of Inkling-Small
Everything that runs on NVIDIA B200
Get the MXFP4_MOE file on Hugging Face
llama-server -hf unsloth/Inkling-Small-GGUF:MXFP4_MOEFiles and context on NVIDIA B200
Each file of Inkling-Small (and of GGUF versions the site has analyzed), and how much context fits next to it, with the context cache in 16-bit or 8-bit.
| File | Size | On NVIDIA B200 | Context (16-bit) | Context (8-bit) |
|---|---|---|---|---|
| As published | 532 GB | Doesn't fit | — | — |
| BF16 | 527 GB | Doesn't fit | — | — |
| UD-Q8_K_XL | 288 GB | Doesn't fit | — | — |
| Q8_0 | 280 GB | Doesn't fit | — | — |
| UD-Q6_K_XL | 240 GB | Doesn't fit | — | — |
| UD-Q6_K | 219 GB | Doesn't fit | — | — |
| UD-Q5_K_XL | 197 GB | Doesn't fit | — | — |
| UD-Q5_K_M | 196 GB | Doesn't fit | — | — |
| UD-Q5_K_S | 184 GB | Doesn't fit | — | — |
| UD-Q4_K_XL | 163 GB | Doesn't fit | — | — |
| UD-Q4_K_M | 163 GB | Doesn't fit | — | — |
| MXFP4_MOE | 158 GB | Tight | 132K | 267K |
| UD-Q4_K_S | 152 GB | Tight | 327K | 656K |
| UD-IQ4_NL | 130 GB | Tight | 1M | 1M |
| UD-IQ4_XS | 127 GB | Fits | 1M | 1M |
| UD-Q3_K_XL | 120 GB | Fits | 1M | 1M |
| UD-Q3_K_M | 119 GB | Fits | 1M | 1M |
| UD-IQ3_S | 107 GB | Fits | 1M | 1M |
| UD-IQ3_XXS | 97.9 GB | Fits | 1M | 1M |
| UD-Q2_K_XL | 87.9 GB | Fits | 1M | 1M |
| UD-IQ2_M | 82.4 GB | Fits | 1M | 1M |
| UD-IQ2_XXS | 82.3 GB | Fits | 1M | 1M |
| UD-IQ1_M | 78.8 GB | Fits | 1M | 1M |
| UD-IQ1_S | 74.8 GB | Fits | 1M | 1M |