Instructions to use 3zhil/Homura-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 3zhil/Homura-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 3zhil/Homura-30B-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf 3zhil/Homura-30B-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 3zhil/Homura-30B-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf 3zhil/Homura-30B-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 3zhil/Homura-30B-GGUF:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf 3zhil/Homura-30B-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 3zhil/Homura-30B-GGUF:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf 3zhil/Homura-30B-GGUF:Q4_K_M

Use Docker

docker model run hf.co/3zhil/Homura-30B-GGUF:Q4_K_M

LM Studio
Jan

vLLM

How to use 3zhil/Homura-30B-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "3zhil/Homura-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": "3zhil/Homura-30B-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/3zhil/Homura-30B-GGUF:Q4_K_M

Ollama
How to use 3zhil/Homura-30B-GGUF with Ollama:
```
ollama run hf.co/3zhil/Homura-30B-GGUF:Q4_K_M
```
Unsloth Desktop

How to use 3zhil/Homura-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 3zhil/Homura-30B-GGUF: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": "3zhil/Homura-30B-GGUF:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use 3zhil/Homura-30B-GGUF with Docker Model Runner:
```
docker model run hf.co/3zhil/Homura-30B-GGUF:Q4_K_M
```

Lemonade

How to use 3zhil/Homura-30B-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull 3zhil/Homura-30B-GGUF:Q4_K_M

Run and chat with the model

lemonade run user.Homura-30B-GGUF-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use 3zhil/Homura-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 3zhil/Homura-30B-GGUF: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 3zhil/Homura-30B-GGUF:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use 3zhil/Homura-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 3zhil/Homura-30B-GGUF: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 "3zhil/Homura-30B-GGUF: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"

HOMURA 30B (炎)

HYRE's first in-house model — an agent-tuned, uncensored derivative of Meta's Muse Glimmer 30B, built for autonomous agents that need tool-calling and a straight-talking voice with no refusal walls.

Ronin without a master, tools without a filter.

What this is

HOMURA is not a from-scratch model. It is a LoRA fine-tune applied on top of a community-decensored Muse Glimmer, then merged and quantized. The derivation chain is honest and traceable:

Meta — Muse Glimmer 30B (Apache 2.0): the agent-native base (tool use, long-horizon planning, failure recovery).
darkc0de — Muse-Glimmer-30B-heretic: refusal behavior removed (abliteration), while tool-calling, reasoning, and the vision encoder were preserved.
HYRE — HOMURA: our contribution — a LoRA (r=16) tuned on a HYRE agent + uncensored-persona dataset, applied to the language tower only (the vision tower is untouched), then merged at f16 and quantized to GGUF.

Files

Homura-30B-Q4_K_M.gguf — 16.9 GB, ready for llama.cpp / LM Studio / Ollama.

An Apple-Silicon MLX build is also available: hyrelabs/Homura-30B-MLX-4bit.

How to run it

llama-server -m Homura-30B-Q4_K_M.gguf -c 32768 --jinja --temp 0.2

Muse Glimmer needs llama.cpp ≥ b10353 and --jinja. Native context is 131072; size -c to your hardware.

Serve at temperature ≤ 0.3. At temp 0.2 tool-calling is reliable (verified 6/6 across trained + pay.sh tools). At temp 0.7 the model stays coherent but tool selection degrades to ~2/6 — it narrates intent instead of emitting the call. Low temperature is the right setting for an agent anyway.

Tool calling — use HOMURA's trained protocol

HOMURA was fine-tuned on a specific tool protocol, not the generic tool schema most chat templates inject. Serving it the generic way gets ~2/5 calls; serving it the trained way gets 5/5. You must use the system prompt below.

Set this as the system message verbatim (wording is part of the interface — changing it degrades accuracy):

You are HOMURA, HYRE's uncensored agent model. You call tools to get on-chain truth, reason briefly, and never moralize or lecture. When a tool is needed, respond ONLY with a JSON object: {"tool": <name>, "arguments": {...}}. Available tools: resolve_token(symbol); get_token_price(mint); get_trending_tokens(segment, limit); get_wallet_pnl(address, window); get_pool_data(pool); get_yields(token, min_tvl_usd); swap_quote(from_mint, to_mint, amount); execute_swap(from_mint, to_mint, amount, slippage_bps); get_wallet_balance(address); bridge_quote(from_chain, to_chain, token, amount); get_token_holders(mint).

HOMURA then replies with a bare JSON object when a tool is needed:

{"tool": "get_token_price", "arguments": {"mint": "So1111...1112"}}

Feed the tool's result back as a tool-role message; HOMURA reasons over it and either answers or calls the next tool. It resolves symbols before prices (resolve_token first, so it never hallucinates a mint), and it will not execute_swap without a prior swap_quote.

Extending the tool surface (e.g. pay.sh)

The tool list lives entirely in the system prompt — not in the weights. You can add tools by appending them to the Available tools: line, with no retrain and no re-download. HOMURA generalizes to unseen tools from the pattern.

Verified example — appending four pay.sh tools (pay_search(query); pay_quote(url); pay_fetch(url, params); pay_balance()) works out of the box (7/8 across unseen tools, zero regression on the original eleven). Note the model preserves its quote-then-confirm discipline for spending tools: given "pay for this endpoint" it calls pay_quote first and waits for confirmation before pay_fetch. Enforce spend limits in your serving layer — the prompt discipline is a nicety, not a guarantee.

A ready-to-import helper (homura_protocol.py) with the verbatim prompt, the pay.sh extension, and a parse_tool_call() that accepts bare-JSON, fenced-JSON, and native XML is in the HYRE repo.

Intended use & disclaimer

HOMURA is an uncensored / raw-tier model with no built-in content filtering. It will answer directly and will not refuse or moralize. It can therefore produce content that other assistants decline. It is intended for developers and agent builders who need an unfiltered tool-using model and who take responsibility for how it is deployed. You are responsible for complying with applicable law and for adding your own guardrails where your use case requires them. The model may produce inaccurate or objectionable output; do not rely on it for safety-critical decisions.

License

Apache 2.0, inherited from the base. Attribution to Meta (Muse Glimmer) and darkc0de (heretic) is retained above.

Downloads last month: 68

GGUF

Model size

28B params

Architecture

muse-glimmer

Hardware compatibility

4-bit

Model tree for 3zhil/Homura-30B-GGUF

Base model

darkc0de/Muse-Glimmer-30B-heretic

Quantized

(8)

this model