Instructions to use guell00/VELUM-Coder with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use guell00/VELUM-Coder 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 guell00/VELUM-Coder:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf guell00/VELUM-Coder:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf guell00/VELUM-Coder:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf guell00/VELUM-Coder: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 guell00/VELUM-Coder:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf guell00/VELUM-Coder: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 guell00/VELUM-Coder:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf guell00/VELUM-Coder:Q4_K_M

Use Docker

docker model run hf.co/guell00/VELUM-Coder:Q4_K_M

LM Studio
Jan
Ollama
How to use guell00/VELUM-Coder with Ollama:
```
ollama run hf.co/guell00/VELUM-Coder:Q4_K_M
```

Unsloth Studio

How to use guell00/VELUM-Coder 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 guell00/VELUM-Coder 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 guell00/VELUM-Coder to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for guell00/VELUM-Coder to start chatting

How to use guell00/VELUM-Coder with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf guell00/VELUM-Coder:Q4_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": "guell00/VELUM-Coder:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use guell00/VELUM-Coder with Docker Model Runner:
```
docker model run hf.co/guell00/VELUM-Coder:Q4_K_M
```

Lemonade

How to use guell00/VELUM-Coder with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull guell00/VELUM-Coder:Q4_K_M

Run and chat with the model

lemonade run user.VELUM-Coder-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use guell00/VELUM-Coder with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf guell00/VELUM-Coder:Q4_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 guell00/VELUM-Coder:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use guell00/VELUM-Coder with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf guell00/VELUM-Coder:Q4_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 "guell00/VELUM-Coder:Q4_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"

VELUM-Coder / README.md

guell00

Update model card after Q2_K

f396b87 verified about 20 hours ago

preview code

Raw

History Blame

4.75 kB

	---
	base_model: ornith-ai/Ornith-1.5-9B
	datasets:
	- guell00/fds
	library_name: gguf
	pipeline_tag: text-generation
	license: mit
	tags:
	- gguf
	- llama.cpp
	- code
	- coding
	- imatrix
	- qwen3.5
	- ornith
	- lora
	---

	# VELUM-Coder

	VELUM-Coder is a coding-focused fine-tune derived from
	[`ornith-ai/Ornith-1.5-9B`](https://huggingface.co/ornith-ai/Ornith-1.5-9B).

	This repository contains GGUF quantizations generated from the completed recovery adapter.

	- Source adapter: `ornith_fds_recovered_ckpt250`
	- Dataset: `guell00/fds`
	- llama.cpp commit: `f280b26983ad0fdb705a0d9ebf0503e76f2899b0`
	- Generated: 2026-08-25 01:31 UTC

	## Quantization policy

	- `Q8_0` is intentionally generated without an importance matrix.
	- Every Q4, Q3, Q2 and IQ1 build is invoked with the same domain-specific imatrix.
	- No Q5 or Q6 files are produced.
	- The extreme `IQ1_M` build keeps the token embedding and output tensors at `Q8_0` while the body is quantized as `IQ1_M`.
	- The imatrix is generated from code/conversation samples from `guell00/fds`.
	- `<think>...</think>` blocks are removed from assistant calibration text so calibration emphasizes direct answers and code.

	## Recommended files

	Best default: `VELUM-Coder-Q4_K_M.gguf`

	Higher quality: `VELUM-Coder-Q8_0.gguf`

	Lower memory: try `VELUM-Coder-IQ3_M.gguf` or `VELUM-Coder-IQ2_M.gguf`.

	Extreme experiment: `VELUM-Coder-IQ1_M.gguf`. This is not expected to preserve Q4-level quality. Embedding/output protection exists only to reduce catastrophic collapse.

	## Files

	\| Status \| File \| Quant \| Imatrix \| Protection \| Notes \|
	\|---\|---\|---\|---\|---\|---\|
	\| ✅ \| `VELUM-Coder-Q8_0.gguf` \| `Q8_0` \| No \| Normal \| Highest-quality quantized build; intentionally no imatrix. \|
	\| ✅ \| `VELUM-Coder-Q4_K_M.gguf` \| `Q4_K_M` \| Yes \| Normal \| Recommended default Q4. \|
	\| ✅ \| `VELUM-Coder-Q4_K_S.gguf` \| `Q4_K_S` \| Yes \| Normal \| Smaller Q4 K-quant. \|
	\| ✅ \| `VELUM-Coder-IQ4_XS.gguf` \| `IQ4_XS` \| Yes \| Normal \| Compact 4-bit I-Quant. \|
	\| ✅ \| `VELUM-Coder-IQ4_NL.gguf` \| `IQ4_NL` \| Yes \| Normal \| Non-linear 4-bit I-Quant. \|
	\| ✅ \| `VELUM-Coder-Q4_1.gguf` \| `Q4_1` \| Yes \| Normal \| Legacy Q4 variant. \|
	\| ✅ \| `VELUM-Coder-Q4_0.gguf` \| `Q4_0` \| Yes \| Normal \| Legacy compact Q4. \|
	\| ✅ \| `VELUM-Coder-Q3_K_L.gguf` \| `Q3_K_L` \| Yes \| Normal \| Largest Q3 K-quant. \|
	\| ✅ \| `VELUM-Coder-Q3_K_M.gguf` \| `Q3_K_M` \| Yes \| Normal \| Balanced Q3 K-quant. \|
	\| ✅ \| `VELUM-Coder-Q3_K_S.gguf` \| `Q3_K_S` \| Yes \| Normal \| Smaller Q3 K-quant. \|
	\| ✅ \| `VELUM-Coder-IQ3_M.gguf` \| `IQ3_M` \| Yes \| Normal \| Medium 3-bit I-Quant. \|
	\| ✅ \| `VELUM-Coder-IQ3_S.gguf` \| `IQ3_S` \| Yes \| Normal \| Small 3-bit I-Quant. \|
	\| ✅ \| `VELUM-Coder-IQ3_XS.gguf` \| `IQ3_XS` \| Yes \| Normal \| Extra-small 3-bit I-Quant. \|
	\| ✅ \| `VELUM-Coder-IQ3_XXS.gguf` \| `IQ3_XXS` \| Yes \| Normal \| Extremely compact 3-bit I-Quant. \|
	\| ✅ \| `VELUM-Coder-Q2_K.gguf` \| `Q2_K` \| Yes \| Normal \| Aggressive Q2 K-quant. \|
	\| ⏳ \| `VELUM-Coder-Q2_K_S.gguf` \| `Q2_K_S` \| Yes \| Normal \| Small Q2 K-quant. \|
	\| ⏳ \| `VELUM-Coder-IQ2_M.gguf` \| `IQ2_M` \| Yes \| Normal \| Higher-quality 2-bit I-Quant. \|
	\| ⏳ \| `VELUM-Coder-IQ2_S.gguf` \| `IQ2_S` \| Yes \| Normal \| 2-bit I-Quant. \|
	\| ⏳ \| `VELUM-Coder-IQ2_XS.gguf` \| `IQ2_XS` \| Yes \| Normal \| Extra-small 2-bit I-Quant. \|
	\| ⏳ \| `VELUM-Coder-IQ2_XXS.gguf` \| `IQ2_XXS` \| Yes \| Normal \| Extremely compact 2-bit I-Quant. \|
	\| ⏳ \| `VELUM-Coder-Q2_0.gguf` \| `Q2_0` \| Yes \| Normal \| 2.25 bpw Q2_0. \|
	\| ⏳ \| `VELUM-Coder-TQ2_0.gguf` \| `TQ2_0` \| Yes \| Normal \| Experimental ternary ~2-bit build. \|
	\| ⏳ \| `VELUM-Coder-IQ1_M.gguf` \| `IQ1_M` \| Yes \| Q8_0 embed/output \| Extreme build. IQ1_M body + Q8_0 embedding/output protection. \|

	## Importance matrix

	The generated importance matrix is uploaded as:

	`imatrix/VELUM-Coder-imatrix.gguf`

	The calibration text itself is not uploaded; it is only an intermediate generated from `guell00/fds`.

	## llama.cpp

	Example:

	~~~bash
	llama-cli -hf guell00/VELUM-Coder:Q4_K_M -c 16384
	~~~

	Server:

	~~~bash
	llama-server -hf guell00/VELUM-Coder:Q4_K_M -c 16384
	~~~

	Manual download:

	~~~bash
	hf download guell00/VELUM-Coder VELUM-Coder-Q4_K_M.gguf --local-dir .
	~~~

	Then:

	~~~bash
	llama-cli -m VELUM-Coder-Q4_K_M.gguf -c 16384
	~~~

	## Adapter

	The final recovery LoRA adapter is also uploaded under `adapter/` when adapter upload is enabled.

	## Notes

	Low-bit quantization is destructive. Q2 and especially IQ1_M are experimental options for constrained hardware, not substitutes for Q4/Q8 quality.

	GGUFs are produced directly by `llama-quantize` and uploaded after successful quantization. No conversational inference is run by this pipeline.

	## Credits

	- Ornith authors
	- llama.cpp / ggml contributors
	- Hugging Face
	- Unsloth
	- `guell00/fds`