- Can I run it locally?
- Inkling
Inkling VRAM requirements: can I run Inkling Small locally?
Inkling Small has 264 billion parameters, 163 GB at 4-bit. Its smallest file, UD-IQ1_S (74.8 GB), is the one that fits a 128 GB machine.
Check your machine
Won't fit 165 GB
Weights, cache and overhead add up to more than the card and the RAM together.
- Weights 163 GB
- Context cache 1.1 GB
- Overhead 0.9 GB
- Fast memory
- With system RAM
What we found
Small is relative here. Inkling Small is a mixture of experts with 264 billion parameters, and the file that gets it onto a DGX Spark or a 128 GB Mac is the smallest one, UD-IQ1_S (74.8 GB). At 4-bit it's a 256 GB Mac.
Its cache stays small for a model this size. 35 of its 42 layers only look back 512 tokens, so the full million-token context adds 30.2 GB.
Inkling is the big one at 947 billion parameters. Its smallest file, UD-IQ1_S (270 GB), fits a 512 GB Mac Studio.
Inkling Small
264 billion parameters · mixture of experts, 6 of 256 active · up to 1M context · 1.2 million downloads in 30 days
Inkling Small needs about 165 GB at UD-Q4_K_M with a 32K-token context. The file itself is 163 GB, and the cache for 32K tokens adds 1.1 GB. No single consumer card holds it. A Mac with 256 GB of unified memory does, and a PC can split it between a GPU and system RAM. It routes each token through 6 of 256 experts, so it runs faster than its size suggests.
35 of its layers only look back 512 tokens, so they stop growing early, and the whole cache at 1M is 30.2 GB.
| Quant | File | At 8K | At 32K | Smallest GPU | Smallest Mac |
|---|---|---|---|---|---|
UD-IQ1_S | 74.8 GB | 76.1 GB | 76.8 GB | 96 GB | 96 GB |
UD-Q2_K_XL | 87.9 GB | 89.2 GB | 89.9 GB | 96 GB | 96 GB* |
UD-Q3_K_XL | 120 GB | 121 GB | 122 GB | none | 128 GB* |
UD-Q4_K_M | 163 GB | 164 GB | 165 GB | none | 256 GB |
UD-Q5_K_XL | 197 GB | 198 GB | 199 GB | none | 256 GB |
UD-Q6_K_XL | 240 GB | 241 GB | 242 GB | none | 256 GB* |
Q8_0 | 280 GB | 282 GB | 282 GB | none | 512 GB |
BF16 | 527 GB | 529 GB | 529 GB | none | 512 GB* |
Smallest GPU and Mac are for the 32K total. * with the Mac's GPU memory limit raised.
Sources: GGUF files · config.json
Inkling
947 billion parameters · mixture of experts, 6 of 256 active · up to 1M context · 588 thousand downloads in 30 days
At UD-Q4_K_XL, the Inkling file is 587 GB, and with 32K tokens of context and runtime overhead it comes to about 590 GB. That's past a 512 GB Mac. The smallest file, UD-IQ1_S at 270 GB, fits a 512 GB Mac. It routes each token through 6 of 256 experts, so it runs faster than its size suggests.
A sliding window of 512 tokens on 55 layers keeps the cache small: 47.7 GB at the full 1M.
| Quant | File | At 8K | At 32K | Smallest GPU | Smallest Mac |
|---|---|---|---|---|---|
UD-IQ1_S | 270 GB | 272 GB | 273 GB | none | 512 GB |
UD-Q2_K_XL | 317 GB | 319 GB | 320 GB | none | 512 GB |
UD-Q3_K_XL | 433 GB | 435 GB | 436 GB | none | 512 GB* |
UD-Q4_K_XL | 587 GB | 589 GB | 590 GB | none | none |
Q8_0 | 857 GB | 858 GB | 860 GB | none | none |
BF16 | 1,894 GB | 1,896 GB | 1,897 GB | none | none |
Smallest GPU and Mac are for the 32K total. * with the Mac's GPU memory limit raised.
Sources: GGUF files · config.json
How the numbers are worked out
Total memory is three parts added together. Weights are the exact size of the GGUF file on Hugging Face, all parts summed when a quant is split. The context cache (the KV cache) comes from each model's own config.json:
cache = layers × kv_heads × head_dim × 2 (K and V) × 2 bytes × tokens
Sliding-window layers stop at their window plus llama.cpp's 512-token batch. Linear-attention and Mamba layers keep a fixed state instead, which we add once. Latent-attention models (DeepSeek, Kimi, GLM-5) cache kv_lora_rank + rope dimensions per layer, the way llama.cpp does. Overhead is our allowance: 0.5 GB for the GPU context plus the output buffer, 512 × vocabulary × 4 bytes.
A "24 GB" card holds 24 GiB, which is 25.8 GB in the decimal units file sizes use, and we count it that way. Macs get the share Metal allows by default, two thirds of unified memory up to 32 GB and three quarters above (llama.cpp discussion 2182); CPU-only machines and the DGX Spark keep 4 GB back for the operating system. All of it assumes llama.cpp's defaults: an f16 cache and one conversation at a time. Quantizing the cache to q8_0 halves it.
Files are Unsloth's GGUF uploads, the most downloaded quantizations on Hugging Face, and the list refreshes daily from Hugging Face's API. A model joins when its repo passes 20,000 downloads in 30 days. Last checked September 25, 2026.