Moe_Scratch needs about 30.0 GB of memory at Q4_K_M (a 28.4 GB file) with an 8K context window, so it fits on a 32 GB GPU. This repository has 21 quants, from 17.3 GB (Q2_K) to 33.2 GB (Q5_K_M). Longer context needs more: at 32K tokens the KV cache alone adds about 4.3 GB.
Moe_Scratch is a 57.6B-parameter Moe model in GGUF format, quantized by RichardErkhov. Its trained context length is 32K tokens. The repository has 21 quant files, from 17.3 GB (Q2_K) to 33.2 GB (Q5_K_M).
Quants & memory
| Quant | File | Est. memory | vs VRAM | Result | |
|---|---|---|---|---|---|
| Q2_K | 17.3 GB | 18.9 GB | Too big | Get file → | |
| IQ3_M | 17.4 GB | 18.9 GB | Too big | Get file → | |
| IQ3_XS | 18.8 GB | 20.4 GB | Too big | Get file → | |
| IQ3_S | 20.4 GB | 22.0 GB | Too big | Get file → | |
| Q3_K_S | 20.4 GB | 22.0 GB | Too big | Get file → | |
| Q3_K | 22.5 GB | 24.1 GB | Too big | Get file → | |
| Q3_K_M | 22.5 GB | 24.1 GB | Too big | Get file → | |
| Q3_K_L | 24.2 GB | 25.7 GB | Too big | Get file → | |
| IQ4_XS | 25.4 GB | 26.9 GB | Too big | Get file → | |
| Q4_0 | 26.4 GB | 28.0 GB | Too big | Get file → | |
| IQ4_NL | 26.7 GB | 28.3 GB | Too big | Get file → | |
| Q4_K_S | 26.7 GB | 28.3 GB | Too big | Get file → | |
| Q4_K | 28.4 GB | 30.0 GB | Too big | Get file → | |
| Q4_K_M | 28.4 GB | 30.0 GB | Too big | Get file → | |
| Q4_1 | 29.3 GB | 30.9 GB | Too big | Get file → | |
| Q5_1 | 29.6 GB | 31.2 GB | Too big | Get file → | |
| Q6_K | 29.6 GB | 31.2 GB | Too big | Get file → | |
| Q5_0 | 32.2 GB | 33.8 GB | Too big | Get file → | |
| Q5_K_S | 32.2 GB | 33.8 GB | Too big | Get file → | |
| Q5_K | 33.2 GB | 34.8 GB | Too big | Get file → | |
| Q5_K_M | 33.2 GB | 34.8 GB | Too big | Get file → |
What GPU do you need for Moe_Scratch?
These figures add the file size, an fp16 KV cache at an 8K context window (worked out from this model’s own header) and about 0.5 GB for the runtime. They are estimates, so leave a little headroom, and expect to use less context or a smaller quant if you are right at the limit.
- 4 GB VRAM: no quant fits at an 8K context.
- 6 GB VRAM: no quant fits at an 8K context.
- 8 GB VRAM: no quant fits at an 8K context.
- 12 GB VRAM: no quant fits at an 8K context.
- 16 GB VRAM: no quant fits at an 8K context.
- 24 GB VRAM: the best quant that fits is IQ3_S (20.4 GB file, about 22.0 GB in use).
- 32 GB VRAM: the best quant that fits is Q6_K (29.6 GB file, about 31.2 GB in use).
- 48 GB VRAM: the best quant that fits is Q5_K (33.2 GB file, about 34.8 GB in use).
Which Moe_Scratch quant should you choose?
Q4_K_M is the usual starting point for Moe_Scratch: it balances file size and quality well. Here it is a 28.4 GB file.
If memory is tight, the smallest file is Q2_K at 17.3 GB, with the quality trade-offs that come with fewer bits. Use it only when memory is very tight. Expect a clear drop in quality, which is usually worse on small models and easier to live with on large ones.
New to quant names? Read GGUF quantization explained.
Specifications
Architecture from the GGUF header
Frequently asked questions about Moe_Scratch
How much VRAM does Moe_Scratch need?
About 30.0 GB at Q4_K_M with an 8K context window. That is the 28.4 GB file, plus the KV cache and roughly 0.5 GB of runtime overhead. Smaller quants need less and longer context needs more.
Can I run Moe_Scratch on an 8 GB GPU?
Not at an 8K context. The smallest file needs about 18.9 GB, so you would need a larger GPU or CPU offloading.
What is the file size of Moe_Scratch GGUF?
It depends on the quant: Q2_K is 17.3 GB, IQ3_M is 17.4 GB, IQ3_XS is 18.8 GB, IQ3_S is 20.4 GB, Q3_K_S is 20.4 GB, Q3_K is 22.5 GB and 15 more.
What context length does Moe_Scratch support?
The GGUF header lists a trained context length of 32,768 tokens (32K). The KV cache grows with the context you use: about 1.1 GB at 8K and 4.3 GB at 32K.
Where can I download Moe_Scratch GGUF?
The files are on Hugging Face in the RichardErkhov/shitshow123_-_moe_scratch-gguf repository. The download buttons on this page open the file on Hugging Face. ggufmodels.com does not host any model files.
What license does Moe_Scratch use?
Hugging Face lists the license as apache-2.0. Licenses can differ between the quantized files and the original model, so check the repository and the base model before commercial use.