Instructions to use deucebucket/Granite-4.1-30B-Cerebellum-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use deucebucket/Granite-4.1-30B-Cerebellum-GGUF with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf deucebucket/Granite-4.1-30B-Cerebellum-GGUF:Q3_K_M
# Run inference directly in the terminal:
llama cli -hf deucebucket/Granite-4.1-30B-Cerebellum-GGUF:Q3_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf deucebucket/Granite-4.1-30B-Cerebellum-GGUF:Q3_K_M
# Run inference directly in the terminal:
llama cli -hf deucebucket/Granite-4.1-30B-Cerebellum-GGUF:Q3_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf deucebucket/Granite-4.1-30B-Cerebellum-GGUF:Q3_K_M
# Run inference directly in the terminal:
./llama-cli -hf deucebucket/Granite-4.1-30B-Cerebellum-GGUF:Q3_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf deucebucket/Granite-4.1-30B-Cerebellum-GGUF:Q3_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf deucebucket/Granite-4.1-30B-Cerebellum-GGUF:Q3_K_M

Use Docker

docker model run hf.co/deucebucket/Granite-4.1-30B-Cerebellum-GGUF:Q3_K_M

LM Studio
Jan
Ollama
How to use deucebucket/Granite-4.1-30B-Cerebellum-GGUF with Ollama:
```
ollama run hf.co/deucebucket/Granite-4.1-30B-Cerebellum-GGUF:Q3_K_M
```
Unsloth Desktop

How to use deucebucket/Granite-4.1-30B-Cerebellum-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf deucebucket/Granite-4.1-30B-Cerebellum-GGUF:Q3_K_M

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "deucebucket/Granite-4.1-30B-Cerebellum-GGUF:Q3_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use deucebucket/Granite-4.1-30B-Cerebellum-GGUF with Docker Model Runner:
```
docker model run hf.co/deucebucket/Granite-4.1-30B-Cerebellum-GGUF:Q3_K_M
```

Lemonade

How to use deucebucket/Granite-4.1-30B-Cerebellum-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull deucebucket/Granite-4.1-30B-Cerebellum-GGUF:Q3_K_M

Run and chat with the model

lemonade run user.Granite-4.1-30B-Cerebellum-GGUF-Q3_K_M

List all available models

lemonade list

Hermes Agent

How to use deucebucket/Granite-4.1-30B-Cerebellum-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf deucebucket/Granite-4.1-30B-Cerebellum-GGUF:Q3_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default deucebucket/Granite-4.1-30B-Cerebellum-GGUF:Q3_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use deucebucket/Granite-4.1-30B-Cerebellum-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf deucebucket/Granite-4.1-30B-Cerebellum-GGUF:Q3_K_M

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "deucebucket/Granite-4.1-30B-Cerebellum-GGUF:Q3_K_M" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Granite 4.1-30B — Cerebellum GGUF

Ablation-guided mixed-precision quantization of ibm-granite/granite-4.1-30b. 30B parameters, dense architecture with GQA, 64 layers.

What is Cerebellum?

Instead of uniform quantization, we measure which weight groups survive aggressive compression and which don't. Groups that tolerate Q2_K get demoted; groups that don't stay at Q3_K_M or higher. The result: smaller files with less quality loss than uniform quants of the same size.

Files

File	Size	Description
`Granite-4.1-30B-Cerebellum-v2-Q3_K_M.gguf`	13 GB	Optimal mix — 3 groups demoted (attn_k, attn_q, attn_output), 4 kept at Q3_K_M
`Granite-4.1-30B-Cerebellum-v1-Q3_K_M.gguf`	12 GB	Aggressive — 5 groups demoted (all attn + ffn_gate)

Benchmarks

Evaluated using our standardized benchmark suite (ARC-Challenge, HellaSwag, MMLU, HumanEval) with temperature=0, no thinking mode.

The model-index metadata in this card's frontmatter mirrors the recommended v2 build, measured with the local llama.cpp benchmark harness on RTX 3090.

Cerebellum v2 (13 GB) — Recommended

Benchmark	Score	Questions
ARC-Challenge	91.6%	1,172
HellaSwag	88.9%	10,042
MMLU	73.5%	11,643
HumanEval	82.3%	164

Size vs Quality

Model	Size	BPW	PPL (wiki)
Q3_K_M (baseline)	14 GB	3.94	8.3736
Cerebellum v2	13 GB	3.76	8.4912
Cerebellum v1	12 GB	3.50	9.1405

v2 saves 1 GB (7%) over Q3_K_M with only +1.4% perplexity increase — and the 3 demoted groups actually improved perplexity individually during ablation.

Methodology

Group ablation: Demote each of 7 weight groups (attn_k, attn_q, attn_v, attn_output, ffn_gate, ffn_up, ffn_down) to Q2_K individually. Measure PPL impact.
Identify improvers: Three groups (attn_k, attn_q, attn_output) showed lower PPL when demoted — the Q3_K_M precision was actually hurting these layers.
Build optimal mix: v2 demotes only the 3 groups that improve; v1 additionally demotes attn_v and ffn_gate.

Ablation Results

Group	PPL when demoted	Delta vs baseline
attn_k	8.1639	-0.2097 (improved!)
attn_q	8.3144	-0.0592 (improved!)
attn_output	8.3539	-0.0197 (improved!)
attn_v	8.4713	+0.0977
ffn_gate	8.5967	+0.2231
ffn_up	9.0587	+0.6851
ffn_down	9.0190	+0.6454

v2 Override Map

Demoted (Q2_K): attn_k, attn_q, attn_output (all 64 layers)
Sacred (kept at Q3_K_M): attn_v, ffn_gate, ffn_up, ffn_down

Usage

Works with any llama.cpp-compatible tool:

# llama.cpp
./llama-server --model Granite-4.1-30B-Cerebellum-v2-Q3_K_M.gguf -ngl 99 --ctx-size 4096

# Ollama (create Modelfile pointing to the GGUF)
# LM Studio (drag and drop)
# koboldcpp, text-generation-webui, etc.

Hardware Requirements

v2 (13 GB): Fits in 16 GB VRAM with room for context. RTX 4060 Ti 16GB, RTX 3090, etc.
v1 (12 GB): Fits in 16 GB VRAM with generous context, or tight in 12 GB.

Credits

Quantized with Cerebellum — ablation-guided mixed-precision quantization by deucebucket.
Base model by IBM Granite.

Downloads last month: 48

GGUF

Model size

29B params

Architecture

granite

Hardware compatibility

3-bit

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for deucebucket/Granite-4.1-30B-Cerebellum-GGUF

Base model

ibm-granite/granite-4.1-30b

Quantized

(44)

this model

Evaluation results

normalized accuracy on AI2 Reasoning Challenge
test set Local benchmark run (RTX 3090, llama.cpp)

0.916
accuracy on HellaSwag
validation set Local benchmark run (RTX 3090, llama.cpp)

0.889
accuracy on MMLU
test set Local benchmark run (RTX 3090, llama.cpp)

0.735
pass@1 on HumanEval (pass@1)
test set Local benchmark run (RTX 3090, llama.cpp)

0.823
perplexity on WikiText-2 Perplexity
test set Local benchmark run (RTX 3090, llama.cpp)

8.491