Instructions to use unsloth/Olmo-3-7B-Instruct-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 unsloth/Olmo-3-7B-Instruct-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 unsloth/Olmo-3-7B-Instruct-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
llama cli -hf unsloth/Olmo-3-7B-Instruct-GGUF:UD-Q4_K_XL

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf unsloth/Olmo-3-7B-Instruct-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
llama cli -hf unsloth/Olmo-3-7B-Instruct-GGUF:UD-Q4_K_XL

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 unsloth/Olmo-3-7B-Instruct-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
./llama-cli -hf unsloth/Olmo-3-7B-Instruct-GGUF:UD-Q4_K_XL

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 unsloth/Olmo-3-7B-Instruct-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
./build/bin/llama-cli -hf unsloth/Olmo-3-7B-Instruct-GGUF:UD-Q4_K_XL

Use Docker

docker model run hf.co/unsloth/Olmo-3-7B-Instruct-GGUF:UD-Q4_K_XL

LM Studio
Jan
Ollama
How to use unsloth/Olmo-3-7B-Instruct-GGUF with Ollama:
```
ollama run hf.co/unsloth/Olmo-3-7B-Instruct-GGUF:UD-Q4_K_XL
```

Unsloth Studio

How to use unsloth/Olmo-3-7B-Instruct-GGUF with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for unsloth/Olmo-3-7B-Instruct-GGUF to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for unsloth/Olmo-3-7B-Instruct-GGUF to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for unsloth/Olmo-3-7B-Instruct-GGUF to start chatting

How to use unsloth/Olmo-3-7B-Instruct-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf unsloth/Olmo-3-7B-Instruct-GGUF:UD-Q4_K_XL

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "unsloth/Olmo-3-7B-Instruct-GGUF:UD-Q4_K_XL"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use unsloth/Olmo-3-7B-Instruct-GGUF with Docker Model Runner:
```
docker model run hf.co/unsloth/Olmo-3-7B-Instruct-GGUF:UD-Q4_K_XL
```

Lemonade

How to use unsloth/Olmo-3-7B-Instruct-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull unsloth/Olmo-3-7B-Instruct-GGUF:UD-Q4_K_XL

Run and chat with the model

lemonade run user.Olmo-3-7B-Instruct-GGUF-UD-Q4_K_XL

List all available models

lemonade list

Hermes Agent

How to use unsloth/Olmo-3-7B-Instruct-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 unsloth/Olmo-3-7B-Instruct-GGUF:UD-Q4_K_XL

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 unsloth/Olmo-3-7B-Instruct-GGUF:UD-Q4_K_XL

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use unsloth/Olmo-3-7B-Instruct-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf unsloth/Olmo-3-7B-Instruct-GGUF:UD-Q4_K_XL

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 "unsloth/Olmo-3-7B-Instruct-GGUF:UD-Q4_K_XL" \
  --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"

Possible incorrect YaRN `attention_factor` metadata in Olmo-3-7B-Instruct GGUF

by John6666 - opened 16 days ago

Discussion

John6666

16 days ago

•

edited 16 days ago

Hi.

I found a reproducible BF16 logit mismatch in the current unsloth/Olmo-3-7B-Instruct-GGUF that appears to be caused by the legacy generic YaRN attention-factor metadata.

The original allenai/Olmo-3-7B-Instruct config contains YaRN scaling with factor = 8.0 and attention_factor = 1.2079441541679836.

Reproduction

Using the same BF16 GGUF, same runtime/environment, and the same three prompts:

GGUF metadata condition	Mean Pearson vs native HF BF16	NRMSE
baseline	0.978020076	0.208896
exact-F32 identity: `olmo2.rope.scaling.attn_factor = 1.2079441547393799`	0.978020076	0.208896
`olmo2.rope.scaling.attn_factor = 1.0` — run 1	0.999812275	0.019375
`olmo2.rope.scaling.attn_factor = 1.0` — run 2	0.999812275	0.019375

Controls were deterministic:

baseline and exact-F32 identity logits were bit-for-bit identical
both 1.0 runs were bit-for-bit identical
changing only this metadata scalar removes almost all of the large native-BF16/GGUF-BF16 mismatch

Relevant upstream evidence

The initial llama.cpp OLMo3 support PR #16015 added the HF-to-GGUF conversion in bd1f3e5. That conversion copied the Hugging Face YaRN attention_factor into the generic GGUF RoPE attention-factor field:

self.gguf_writer.add_rope_scaling_attn_factors(
    rope_scaling["attention_factor"]
)

The same PR's BF16 logits verification reported NMSE = 2.331092e-03 and classified it as PASS. This therefore does not appear to rule out the mismatch observed above.

Later, llama.cpp PR #18013, merged as 5c8a717, refactored RoPE conversion into common handling. The OLMo-specific generic attention_factor mapping was removed, and current YaRN conversion uses the YaRN-specific metadata handling instead.

Suggested fix / check

Could you regenerate the BF16 GGUF directly from the original HF model using the current llama.cpp converter and compare its logits against native HF BF16?

If the fresh BF16 conversion confirms the result, the existing quantized files derived from the older GGUF may also need regeneration.

The important part is to regenerate from the original HF weights rather than repairing or re-quantizing the existing GGUF, so the legacy metadata mapping is not inherited.

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment