Instructions to use jaimef21/crystal-qwen-v3-30b-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 jaimef21/crystal-qwen-v3-30b-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 jaimef21/crystal-qwen-v3-30b-gguf
# Run inference directly in the terminal:
llama cli -hf jaimef21/crystal-qwen-v3-30b-gguf

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf jaimef21/crystal-qwen-v3-30b-gguf
# Run inference directly in the terminal:
llama cli -hf jaimef21/crystal-qwen-v3-30b-gguf

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 jaimef21/crystal-qwen-v3-30b-gguf
# Run inference directly in the terminal:
./llama-cli -hf jaimef21/crystal-qwen-v3-30b-gguf

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 jaimef21/crystal-qwen-v3-30b-gguf
# Run inference directly in the terminal:
./build/bin/llama-cli -hf jaimef21/crystal-qwen-v3-30b-gguf

Use Docker

docker model run hf.co/jaimef21/crystal-qwen-v3-30b-gguf

LM Studio
Jan

vLLM

How to use jaimef21/crystal-qwen-v3-30b-gguf with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "jaimef21/crystal-qwen-v3-30b-gguf"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jaimef21/crystal-qwen-v3-30b-gguf",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/jaimef21/crystal-qwen-v3-30b-gguf

Ollama
How to use jaimef21/crystal-qwen-v3-30b-gguf with Ollama:
```
ollama run hf.co/jaimef21/crystal-qwen-v3-30b-gguf
```
Unsloth Desktop

How to use jaimef21/crystal-qwen-v3-30b-gguf with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf jaimef21/crystal-qwen-v3-30b-gguf

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": "jaimef21/crystal-qwen-v3-30b-gguf"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use jaimef21/crystal-qwen-v3-30b-gguf with Docker Model Runner:
```
docker model run hf.co/jaimef21/crystal-qwen-v3-30b-gguf
```

Lemonade

How to use jaimef21/crystal-qwen-v3-30b-gguf with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull jaimef21/crystal-qwen-v3-30b-gguf

Run and chat with the model

lemonade run user.crystal-qwen-v3-30b-gguf-{{QUANT_TAG}}

List all available models

lemonade list

Hermes Agent

How to use jaimef21/crystal-qwen-v3-30b-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 jaimef21/crystal-qwen-v3-30b-gguf

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 jaimef21/crystal-qwen-v3-30b-gguf

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use jaimef21/crystal-qwen-v3-30b-gguf with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf jaimef21/crystal-qwen-v3-30b-gguf

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 "jaimef21/crystal-qwen-v3-30b-gguf" \
  --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"

crystal-qwen-v3-30b-gguf

LoRA-fine-tuned Qwen3-Coder-30B-A3B-Instruct specialised for the Crystal programming language. Quantized to Q8_0 GGUF for local inference via Ollama or llama.cpp.

This is v3 — a full-pipeline retrain on a substantially larger and cleaner corpus than v2 (crystal-qwen3.6-30b-gguf).

Quick start (Ollama)

# Download the GGUF + Modelfile from this repo, then:
ollama create crystal-qwen-v3 -f Modelfile
ollama run crystal-qwen-v3 "Write a Crystal class Item with name and quantity that serializes to JSON."

What v3 changes vs v2

	v2 (`3.6`)	v3
CPT tokens	~3 M	~50 M (top-500 Crystal GitHub repos + stdlib + book + RFCs)
SFT pairs	~600	~3 000 (mined + Claude-Haiku-4.5 augmented, compile-gated)
DPO pairs	37	74 (chosen = idiomatic Crystal; rejected = Ruby-style or wrong)
LoRA rank / α	32 / 64	64 / 128
Quantization	Q4_K_M	Q8_0
Targets	attn only	attn + MLP + MoE `experts.gate_up_proj/down_proj`

Eval

Held-out eval (30 natural-language Crystal tasks; idiom score + crystal build --no-codegen compile gate). Higher is better.

Model	Idiom	Compile pass	Total
crystal-qwen-v3 (this model)	+76	26/28 (93 %)	+206
jaimef21/crystal-qwen3.6-30b (v2)	+67	21/28 (75 %)	+172
Qwen3-Coder-30B-A3B-Instruct (base)	+71	21/28 (75 %)	+176

v3 beats both prior baselines on every axis. v2 actually slightly trailed the base model on idiom (+67 vs +71) — v3 is the first checkpoint of this lineage that is unambiguously better than vanilla Qwen3-Coder at Crystal.

Caveat: v3 ships at Q8_0 while v2 and base measurements above are Q4_K_M, so some compile-gate fidelity gap is attributable to less quantization noise.

Files

crystal-qwen-v3-30b.gguf — 32 GB, Q8_0 (8.51 BPW)
Modelfile — Ollama Modelfile (chatml template + Crystal system prompt)
README.md — this file

Reproducing

Full reproduction pipeline (data scraping → training on RunPod H200 → GGUF quantization → eval) is documented at: github.com/jaimef/crystal-lora — see REPRODUCE.md.

License

Inherits Qwen's Apache 2.0 license from the base model.

Downloads last month: 6

GGUF

Model size

31B params

Architecture

qwen3moe

Hardware compatibility

We're not able to determine the quantization variants.

View all variants

Model tree for jaimef21/crystal-qwen-v3-30b-gguf

Base model

Qwen/Qwen3-Coder-30B-A3B-Instruct

Adapter

(54)

this model