grug 27b mtp needs about 19.5 GB of memory at Q4_K_M (a 16.8 GB file) with an 8K context window, so it fits on a 24 GB GPU. This repository has 3 quants, from 16.8 GB (Q4_K_M) to 29.0 GB (Q8_0). Longer context needs more: at 32K tokens the KV cache alone adds about 8.7 GB.
Grug 27B MTP – 27‑billion‑parameter Qwen3.5 model with multi‑token prediction head
Grug 27B MTP is a 27‑billion‑parameter Qwen3.5 model with a trained multi‑token prediction (MTP) head for speculative decoding. The backbone is identical to grug‑27b, but the MTP head is retrained on grug v2.1 data. It is released under Apache‑2.0 and supports 262,144‑token context.
Good at
- Fast speculative decoding with MTP head
- Large context support up to 262,144 tokens
- Stable performance on Qwen3.5 architecture
Limits
- Requires llama.cpp build with qwen3_5 MTP support
- Large VRAM needed for full precision
- No vision or audio capabilities
Best for
- High‑throughput inference with speculative decoding
- Long‑form content generation
- Research on multi‑token prediction models
Who it’s for. Developers and researchers needing a large Qwen3.5 model with speculative decoding. The model is suitable for environments where 8‑k context inference is common and where the MTP head can be leveraged for speedup.
Quants & memory
| Quant | File | Est. memory | vs VRAM | Result | |
|---|---|---|---|---|---|
| Q4_K_M | 16.8 GB | 19.5 GB | Too big | Get file → | |
| Q5_K_M | 19.5 GB | 22.2 GB | Too big | Get file → | |
| Q8_0 | 29.0 GB | 31.7 GB | Too big | Get file → |
For 8 GB VRAM, use Q4_K_M (16.81 GB file, ~19.5 GB VRAM at 8 k context). For 12 GB VRAM, Q5_K_M (19.54 GB file, ~22.2 GB VRAM). Full precision Q8_0 requires 29.05 GB file and ~31.7 GB VRAM.
What GPU do you need for grug 27b mtp?
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 Q5_K_M (19.5 GB file, about 22.2 GB in use).
- 32 GB VRAM: the best quant that fits is Q8_0 (29.0 GB file, about 31.7 GB in use).
Which grug 27b mtp quant should you choose?
Q4_K_M is the usual starting point for grug 27b mtp: it balances file size and quality well. Here it is a 16.8 GB file.
The Q8_0 file is 29.0 GB, about 1.7 times the size of Q4_K_M. It keeps the weights closer to the original, which only matters if you have the memory to spare.
New to quant names? Read GGUF quantization explained.
Specifications
Architecture from the GGUF header
Source
Frequently asked questions about grug 27b mtp
How much VRAM does grug 27b mtp need?
About 19.5 GB at Q4_K_M with an 8K context window. That is the 16.8 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 grug 27b mtp on an 8 GB GPU?
Not at an 8K context. The smallest file needs about 19.5 GB, so you would need a larger GPU or CPU offloading.
What is the file size of grug 27b mtp GGUF?
It depends on the quant: Q4_K_M is 16.8 GB, Q5_K_M is 19.5 GB, Q8_0 is 29.0 GB.
What context length does grug 27b mtp support?
The GGUF header lists a trained context length of 262,144 tokens (256K). The KV cache grows with the context you use: about 2.2 GB at 8K and 8.7 GB at 32K.
Where can I download grug 27b mtp GGUF?
The files are on Hugging Face in the ProCreations/grug-27b-mtp-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 grug 27b mtp 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.
Can I run Grug 27B MTP on 8 GB VRAM?
Use the Q4_K_M quantized file; it needs about 19.5 GB VRAM at 8 k context, so 8 GB is insufficient.
Does the model support vision or audio?
No, the model has no vision or audio capabilities.
What is the context length of this model?
It supports up to 262,144 tokens of context.