---
license: apache-2.0
base_model: ProCreations/grug-3b
tags:
- grug
- gguf
- llama.cpp
- reasoning
- token-efficient
language:
- en
pipeline_tag: text-generation
---

# grug-3b-qat-q4-gguf

q4 that survive the squeeze.

normal q4 round the weight after training and hope. this one train WITH the
rounding: every linear weight fake-quantized to asymmetric int4 (group 32) on
each forward, straight-through gradient update the bf16 weight underneath. model
learn weight that still work after Q4_K_M round them. same recipe as grug-9b-qat
and grug-27b-qat.

trained on same data as [ProCreations/grug-3b](https://huggingface.co/ProCreations/grug-3b),
so grug dialect and adaptive think length come through intact.

| file | size | note |
|---|---|---|
| grug-3b-qat-q4-Q4_K_M.gguf | 2.57 GB | **the point of this repo** |
| grug-3b-qat-q4-f16.gguf | 8.34 GB | qat weights unquantized, roll your own quant |

use the Q4_K_M one. plain (non-qat) quants live
[here](https://huggingface.co/ProCreations/grug-3b-gguf).

## llama.cpp support

Nanbeige4.2 not in upstream llama.cpp yet (issue
[#26086](https://github.com/ggml-org/llama.cpp/issues/26086)). Nanbeige team PR
[#25994](https://github.com/ggml-org/llama.cpp/pull/25994) add it - weight-shared
depth loop, `num_loops=2`. until merge, build from that branch:

```bash
git clone --depth 1 --branch nanbeige42 https://github.com/Nanbeige/llama.cpp
cd llama.cpp && cmake -B build -DCMAKE_BUILD_TYPE=Release && cmake --build build -j
./build/bin/llama-cli -m grug-3b-Q4_K_M.gguf -p "What is 12 times 12?"
```

these gguf converted and load-probed with that branch.