Instructions to use Omega-Factory/gemma-4-31b-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 Omega-Factory/gemma-4-31b-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 Omega-Factory/gemma-4-31b-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Omega-Factory/gemma-4-31b-GGUF:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Omega-Factory/gemma-4-31b-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Omega-Factory/gemma-4-31b-GGUF:Q4_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 Omega-Factory/gemma-4-31b-GGUF:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf Omega-Factory/gemma-4-31b-GGUF:Q4_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 Omega-Factory/gemma-4-31b-GGUF:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Omega-Factory/gemma-4-31b-GGUF:Q4_K_M

Use Docker

docker model run hf.co/Omega-Factory/gemma-4-31b-GGUF:Q4_K_M

LM Studio
Jan

vLLM

How to use Omega-Factory/gemma-4-31b-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Omega-Factory/gemma-4-31b-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Omega-Factory/gemma-4-31b-GGUF",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/Omega-Factory/gemma-4-31b-GGUF:Q4_K_M

Ollama
How to use Omega-Factory/gemma-4-31b-GGUF with Ollama:
```
ollama run hf.co/Omega-Factory/gemma-4-31b-GGUF:Q4_K_M
```
Unsloth Desktop
Docker Model Runner
How to use Omega-Factory/gemma-4-31b-GGUF with Docker Model Runner:
```
docker model run hf.co/Omega-Factory/gemma-4-31b-GGUF:Q4_K_M
```

Lemonade

How to use Omega-Factory/gemma-4-31b-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Omega-Factory/gemma-4-31b-GGUF:Q4_K_M

Run and chat with the model

lemonade run user.gemma-4-31b-GGUF-Q4_K_M

List all available models

lemonade list

Atomic Chat

Gemma 4 31B Dense GGUF Precision Ladder

Official Release & Multi-Quant Benchmarks:
View full precision comparison tables (Q8_0, Q4_K_M, Q2_K vs. FP16 Base), item-level McNemar discordance flips, and verified cryptographic receipts on the Omega Factory Gemma 4 31B Dense Model Page.

Certified GGUF precision quantization ladder for google/gemma-4-31b dense architecture.

Certified quant ladder

Every rung below was certified on its own exact bytes (sha256-bound). A verdict is never transferred between formats or rungs, and a rung that did not clear its bands is published with its real verdict rather than omitted.

Rung	Verdict	Seeds	Size	SHA256	Receipt
`Q8_0`	CERTIFIED	1	32.64 GB	`22a8ab2d7a3d`	`receipts/cert_Q8_0.json`
`Q4_K_M`	CERTIFIED	1	18.69 GB	`7584d5b09832`	`receipts/cert_Q4_K_M.json`
`Q2_K`	CERTIFIED	1	11.92 GB	`e8e9bcfe6e1a`	`receipts/cert_Q2_K.json`

SCREEN means the result did not use the full seed set and therefore is not a certification.

FP16 Baseline vs. Quantization Precision Comparison

To evaluate quantization precision loss and capability retention, quant variants were benchmarked against the unquantized base model (google/gemma-4-31b) under controlled deterministic evaluation settings.

Precision & Accuracy Matrix vs. FP16 Baseline

Variant / Format	Size (Disk/VRAM)	GSM8K Accuracy	MATH-500 Accuracy	ARC-MC Accuracy	Babble Rate	Status
FP16 Baseline	61.4 GB	86.0% (172/200)	36.5% (73/200)	85.0% (170/200)	0.0333	Baseline Ground Truth
`Q8_0`	32.64 GB	84.5% (Δ -1.5 pp)	37.0% (Δ +0.5 pp)	95.0% (Δ +10.0 pp)	0.0000	CERTIFIED
`Q4_K_M`	18.69 GB	87.5% (Δ +1.5 pp)	35.5% (Δ -1.0 pp)	94.5% (Δ +9.5 pp)	0.0000	CERTIFIED
`Q2_K`	11.92 GB	2.0% (Δ -84.0 pp)	6.0% (Δ -30.5 pp)	91.0% (Δ +6.0 pp)	0.0000	CERTIFIED

Detailed Item-Level Contingency & Flips vs. FP16

Per-item McNemar discordance analysis measuring questions that changed answer status relative to the unquantized FP16 baseline:

Quantization Variant	Category	Quant Correct	FP16 Base Correct	Flips Broke (FP16 ✓ → Quant ✗)	Flips Fixed (FP16 ✗ → Quant ✓)	Net Delta (pp)
`Q8_0`	ARC-MC	190 / 200	170 / 200	0	20	+10.0 pp
`Q8_0`	GSM8K	169 / 200	172 / 200	4	1	-1.5 pp
`Q8_0`	MATH-500	74 / 200	73 / 200	0	1	+0.5 pp
`Q4_K_M`	ARC-MC	189 / 200	170 / 200	0	19	+9.5 pp
`Q4_K_M`	GSM8K	175 / 200	172 / 200	8	11	+1.5 pp
`Q4_K_M`	MATH-500	71 / 200	73 / 200	12	10	-1.0 pp
`Q2_K`	ARC-MC	182 / 200	170 / 200	9	21	+6.0 pp
`Q2_K`	GSM8K	4 / 200	172 / 200	168	0	-84.0 pp
`Q2_K`	MATH-500	12 / 200	73 / 200	65	4	-30.5 pp

Hardware & VRAM Allocation Recommendations

Artifact size is disk usage, not a measured VRAM requirement. Runtime memory and the KV cache depend on architecture, backend, batch settings, and context length. Validate actual deployment with peak-memory profiling.

Reference capacity bands only:

Artifact size (disk)	Reference capacity band
Up to 8 GB	6–8 GB
Over 8 GB through 16 GB	12–16 GB
Over 16 GB through 24 GB	20–24 GB
Over 24 GB through 48 GB	32–48 GB
Over 48 GB	Multi-GPU or CPU offload; no single listed tier

Quant Variant	Artifact Size (disk)	Deployment sizing notes
`Q8_0`	32.64 GB	32–48 GB reference capacity. Candidate only, not a measured fit guarantee.
`Q4_K_M`	18.69 GB	20–24 GB reference capacity. Candidate only, not a measured fit guarantee.
`Q2_K`	11.92 GB	12–16 GB reference capacity. Candidate only, not a measured fit guarantee.

Recommended Sampling Parameters

Thinking Mode:
- temperature = 1.0
- top_p = 0.95
- top_k = 20
- repetition_penalty = 1.0
Instruct Mode:
- temperature = 0.7
- top_p = 0.8
- top_k = 20
- presence_penalty = 1.5

Model Architecture Specifications

Parameters: 32.7 Billion
Layers: 60
Context Length: 262144
Attention: Standard Attention with SWA

Chat template

{{ bos_token }}{% for message in messages %}{% if message['role'] == 'user' %}{{ '<start_of_turn>user
' + message['content'] + '<end_of_turn>
' }}{% elif message['role'] == 'assistant' %}{{ '<start_of_turn>model
' + message['content'] + '<end_of_turn>
' }}{% endif %}{% endfor %}

License and attribution

License: Gemma Terms of Use (inherited from the base model).
Base model: google/gemma-4-31b -- all credit for the underlying model belongs to its authors.
This repository publishes a compressed derivative: experts were pruned and the result re-quantised. It is not a new base model.
Research / free use only. This model inherits a community license: it is published for research and free use, and this repository makes no commercial-serving claims.
You remain responsible for complying with the base model's license terms.

Contact / reporting an issue

Found a problem with these weights or this card? Open a discussion on this repository, or report a security concern to the maintainers via the organisation page at https://huggingface.co/Omega-Factory. We publish corrections and, where warranted, retract a release rather than leave a wrong card standing.

Official Model Page: Gemma 4 31B Dense on Omega Factory
Organization Website: https://omega-factory.ai

Downloads last month: 267

GGUF

Model size

31B params

Architecture

gemma4

Hardware compatibility

2-bit

4-bit

8-bit