---
license: apache-2.0
base_model: ProCreations/grug-35b-v2
tags:
- grug
- gguf
- llama.cpp
- reasoning
- moe
language:
- en
---

# grug-35b-v2-gguf

35b brother reborn, squeezed into rocks. MoE brain: fast like small model,
smart like big model. grug think dense inside `<think>`, answer normal english.
full story + all benchmark: [grug-35b-v2](https://huggingface.co/ProCreations/grug-35b-v2).

## rocks

| file | quant |
|---|---|
| grug-35b-v2-Q8_0.gguf | Q8_0 |
| grug-35b-v2-Q6_K.gguf | Q6_K |
| grug-35b-v2-Q5_K_M.gguf | Q5_K_M |
| grug-35b-v2-Q4_K_M.gguf | Q4_K_M - grug pick this |
| grug-35b-v2-Q3_K_M.gguf | Q3_K_M |
| mmproj-grug-35b-v2-f16.gguf | eye rock (vision) |

every rock load-tested before upload. old grug-35b repeat-word sickness:
extinct in v2 (0 loops in every gauntlet, 100% think closed).


## if rock act broken

single-token spam ("/" forever etc) = NOT the rock. hybrid DeltaNet brain
CANNOT survive llama.cpp context-shift: old builds shift on context overflow
and corrupt the recurrent state into token spam. fix:
- use RECENT llama.cpp (qwen3_5 support; new builds refuse instead of shift)
- agent frontends (OpenCode etc): set `-c 16384` or bigger
- still broken? re-download rock (verify size) + check backend
grug re-test rock after every report: loads clean, zero spam at proper config.

## how run

```bash
llama-server -m grug-35b-v2-Q4_K_M.gguf --mmproj mmproj-grug-35b-v2-f16.gguf \
  -c 16384 --temp 0.6 --top-p 0.95 --top-k 20
```

need recent llama.cpp (qwen3_5_moe arch). MTP not included.
grug made by ProCreations. base brain by DeepReinforce.