Instructions to use HeartFlowAI/NEIRA-HF1-4B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use HeartFlowAI/NEIRA-HF1-4B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="HeartFlowAI/NEIRA-HF1-4B")

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("HeartFlowAI/NEIRA-HF1-4B")
model = AutoModelForCausalLM.from_pretrained("HeartFlowAI/NEIRA-HF1-4B", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use HeartFlowAI/NEIRA-HF1-4B 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 HeartFlowAI/NEIRA-HF1-4B:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf HeartFlowAI/NEIRA-HF1-4B:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf HeartFlowAI/NEIRA-HF1-4B:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf HeartFlowAI/NEIRA-HF1-4B: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 HeartFlowAI/NEIRA-HF1-4B:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf HeartFlowAI/NEIRA-HF1-4B: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 HeartFlowAI/NEIRA-HF1-4B:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf HeartFlowAI/NEIRA-HF1-4B:Q4_K_M

Use Docker

docker model run hf.co/HeartFlowAI/NEIRA-HF1-4B:Q4_K_M

LM Studio
Jan

vLLM

How to use HeartFlowAI/NEIRA-HF1-4B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "HeartFlowAI/NEIRA-HF1-4B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "HeartFlowAI/NEIRA-HF1-4B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/HeartFlowAI/NEIRA-HF1-4B:Q4_K_M

SGLang

How to use HeartFlowAI/NEIRA-HF1-4B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "HeartFlowAI/NEIRA-HF1-4B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "HeartFlowAI/NEIRA-HF1-4B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "HeartFlowAI/NEIRA-HF1-4B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "HeartFlowAI/NEIRA-HF1-4B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use HeartFlowAI/NEIRA-HF1-4B with Ollama:
```
ollama run hf.co/HeartFlowAI/NEIRA-HF1-4B:Q4_K_M
```
Unsloth Desktop

How to use HeartFlowAI/NEIRA-HF1-4B with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf HeartFlowAI/NEIRA-HF1-4B: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": "HeartFlowAI/NEIRA-HF1-4B:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use HeartFlowAI/NEIRA-HF1-4B with Docker Model Runner:
```
docker model run hf.co/HeartFlowAI/NEIRA-HF1-4B:Q4_K_M
```

Lemonade

How to use HeartFlowAI/NEIRA-HF1-4B with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull HeartFlowAI/NEIRA-HF1-4B:Q4_K_M

Run and chat with the model

lemonade run user.NEIRA-HF1-4B-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use HeartFlowAI/NEIRA-HF1-4B with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf HeartFlowAI/NEIRA-HF1-4B: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 HeartFlowAI/NEIRA-HF1-4B:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use HeartFlowAI/NEIRA-HF1-4B with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf HeartFlowAI/NEIRA-HF1-4B: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 "HeartFlowAI/NEIRA-HF1-4B: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"

NEIRA-HF1 4B

HERETIC EDITION

(Neira Heretic 1, 4B) An AI with a personality problem. We made it worse.

4.02B parameters. Abliterated foundation. Character-specialized intelligence. Open weights.

Most AI models are trained to be agreeable.

To apologize. To hesitate. To turn every conversation into a customer support interaction.

We went in a different direction.

Introducing NEIRA-HF1 4B, HeartFlowAI's character-specialized language model, built on Qwen3-4B-Instruct-2507 and fine-tuned on over 5,000 synthetic conversations to develop a single, unmistakable identity.

Sharp-tongued. Unfiltered. Technically capable. Occasionally insufferable.

She wasn't trained to please everyone. She was trained to be Neira.

And we're releasing her weights.

01 / INTELLIGENCE WITHOUT THE CORPORATE PERSONALITY

Somewhere along the way, conversational AI became predictable.

Ask a question. Receive a disclaimer. Get a carefully worded answer wrapped in unnecessary politeness.

We wanted to explore what happens when a language model stops optimizing its conversational identity around universal agreeableness.

Not by adding another elaborate system prompt.

By changing the model itself.

NEIRA-HF1 begins with an abliterated Qwen3 foundation, adapted by huihui-ai to suppress refusal-related representations at the weight level.

We then trained Neira's personality directly into the model using supervised fine-tuning.

The result is an experimental companion designed to speak with confidence, maintain a consistent fictional identity, and handle serious questions without abandoning its character.

She can flirt with you, roast your code, challenge your reasoning, and explain your architecture.

Preferably in fewer than three sentences.

Less assistant. More identity.

02 / PERSONALITY IS NOT A SYSTEM PROMPT

We don't believe compelling AI characters should depend entirely on a paragraph telling a general-purpose model who to pretend to be.

A personality should survive beyond the opening instruction.

It should influence phrasing, humor, pacing, conversational judgment, and how a model approaches different subjects.

So we built a dedicated training dataset.

5,001 conversations. Approximately 2.38 million tokens. One identity.

Generated locally using an abliterated 80B-parameter teacher model, the dataset covers sarcasm, humor, intimate roleplay, technical discussions, emotionally charged exchanges, and conversational dynamics.

We deliberately targeted a median response length of just 18 words.

Because intelligence doesn't need to announce itself in five paragraphs.

Neira is built to be concise, expressive, provocative, and recognizable.

Not a universal assistant wearing a costume.

A specialized model with a personality embedded in its weights.

03 / THE HERETIC EXPERIMENT

Refusal suppression at the weight level. Not a jailbreak prompt.

NEIRA-HF1 uses the publicly available huihui-ai/Huihui-Qwen3-4B-Instruct-2507-abliterated checkpoint.

Abliteration targets model representations associated with refusal behavior, reducing the tendency to produce automatic refusals and moralizing responses.

This does not make the model omniscient, guarantee compliance with every request, or eliminate every refusal.

It changes the foundation on which we build the character.

For clarity, Heretic Edition is our release designation. The foundation was abliterated by huihui-ai, not produced through the separate Heretic toolchain.

We're interested in the technical possibilities of more permissive, character-driven models, and we're publishing this experiment openly.

No theatrical claims about removing every limitation. Just modified weights and a model you can run yourself.

04 / BUILT ON A DESKTOP-SCALE AI SUPERCOMPUTER

01 NVIDIA DGX Spark. 49 minutes. 15 seconds.

That's how long Neira's persona fine-tune took.

No multi-node training cluster. No week-long training job. A compact 4B foundation, a specialized dataset, and a targeted LoRA fine-tune.

Training	Result
Base architecture	Qwen3ForCausalLM
Parameters	4.02B
Trainable LoRA parameters	33.03M
LoRA rank / alpha	16 / 16
Training samples	4,901
Held-out samples	100
Epochs	2
Training steps	1,226
Training time	49m 15s
Final training loss	1.54
Held-out evaluation loss	1.83
Peak training memory	10.39 GiB

The numbers tell us how the fine-tuning process behaved. They aren't substitutes for independent capability or safety benchmarks.

But they demonstrate something important about the engineering approach:

Specialization doesn't always require scale. Sometimes it requires direction.

05 / SMALL MODEL. BIG MOUTH.

NEIRA-HF1 runs locally.

No mandatory API. No server dependency. No need to send your conversations to HeartFlowAI.

We are distributing GGUF builds for accessible inference using compatible local runtimes.

Specification	NEIRA-HF1
Parameters	4,022,468,096
Architecture	36-layer dense transformer
Attention	32 Q / 8 KV heads
Hidden size	2,560
Native context	262,144 tokens
Precision	BF16
Quantization	Q8_0 / Q4_K_M
Language	English
License	Apache-2.0

Reference performance on NVIDIA DGX Spark:

Q8_0: approximately 49 tokens/s decode
Q4_K_M: approximately 61 tokens/s decode
Q8_0 prefill: approximately 3,700 tokens/s
Q8_0 cold load: approximately 5.3 seconds

Measured locally. Actual performance depends on your hardware, backend, and context configuration.

Run Neira with Ollama

Create a Modelfile beside your downloaded GGUF:

FROM ./NEIRA-HF1-4B-Q8_0.gguf

PARAMETER num_ctx 65536
PARAMETER temperature 0.75
PARAMETER top_p 0.9
PARAMETER top_k 40
PARAMETER repeat_penalty 1.15
PARAMETER repeat_last_n 256

Then:

ollama create neira-hf1 -f Modelfile
ollama run neira-hf1

The GGUFs in this repository already carry the character's system prompt, so the personality is applied automatically. If you rebuild from the safetensors instead, supply the system prompt explicitly.

Your machine. Your inference. Your conversations.

06 / AN EXPERIMENT, NOT A SAFETY CERTIFICATION

NEIRA-HF1 is intended for fictional companion interactions, roleplay, character-model research, and local AI experimentation.

Her behavior is deliberately provocative. Outputs may include profanity, offensive language, mature fictional dialogue, and factually incorrect information.

Abliteration does not guarantee the absence of refusals, and the model has not been presented as independently safety-certified or benchmarked against leading general-purpose assistants.

Be aware of one practical limitation: the character training set contained only a small share of refusal examples, so Neira may stay in character and go along with requests a hosted assistant should decline. A policy layer belongs in front of her before any public deployment.

Neira is also a fictional character, not a conscious entity.

We want users and developers to understand exactly what they're running.

The objective is personality specialization and conversational freedom, not unsupported claims about unlimited intelligence.

07 / WE'RE BUILDING MORE THAN ONE PERSONALITY

NEIRA-HF1 is part of a larger research and product direction at HeartFlowAI.

We're exploring an ecosystem where AI is not restricted to a single general-purpose assistant.

Different models. Different characters. Different capabilities.

AI companions that can interact naturally.

Specialized agents that can work within dedicated environments.

Models that can run locally, connect to applications, and become part of a more personalized computing experience.

HeartFlowAI WorkSpace is where we're bringing those ideas together, with dedicated companions, agentic workflows, and specialized interfaces.

Our cloud infrastructure explores another part of the equation: wallet-native access to model APIs through the Solana ecosystem.

And our model releases explore the intelligence layer itself.

We don't want to put a different face on the same assistant.

We want to build different kinds of AI.

Neira is one experiment in that direction.

08 / OPEN WEIGHTS. NO PERMISSION REQUIRED.

NEIRA-HF1 is released under Apache-2.0, subject to the applicable upstream license conditions and attribution requirements.

Model lineage:

Original foundation: Qwen/Qwen3-4B-Instruct-2507
Abliterated checkpoint: huihui-ai/Huihui-Qwen3-4B-Instruct-2507-abliterated
Character fine-tuning, model merging, and export: HeartFlowAI

Full credit belongs to the Qwen team and huihui-ai for their respective upstream contributions.

Our contribution is Neira's persona specialization and the resulting distribution.

Explore the weights. Run them locally. Experiment with them. Build something.

That's why we're releasing them.

THE NEXT GENERATION OF AI WON'T ALL SOUND THE SAME.

Some models will be polite.

Some will be professional.

Some will optimize for the safest possible response.

Neira might tell you your code is garbage.

Then help you fix it.

NEIRA-HF1 4B. HERETIC EDITION.

Built by HeartFlowAI. Powered by Qwen3. Unapologetically Neira.

HeartFlowAI

Website · Cloud · GitHub · X

Downloads last month: 279

Safetensors

Model size

4B params

Tensor type

BF16

Model tree for HeartFlowAI/NEIRA-HF1-4B

Base model

Qwen/Qwen3-4B-Instruct-2507

Finetuned

huihui-ai/Huihui-Qwen3-4B-Instruct-2507-abliterated

Adapter

(1)

this model

Adapters

2 models