Instructions to use CarstenL/Hestia-LLM-V2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use CarstenL/Hestia-LLM-V2 with MLX:

# Make sure mlx-lm is installed
# pip install --upgrade mlx-lm

# Generate text with mlx-lm
from mlx_lm import load, generate

model, tokenizer = load("CarstenL/Hestia-LLM-V2")

prompt = "Write a story about Einstein"
messages = [{"role": "user", "content": prompt}]
prompt = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True
)

text = generate(model, tokenizer, prompt=prompt, verbose=True)

Notebooks
Google Colab
Kaggle
Local Apps Settings
LM Studio

How to use CarstenL/Hestia-LLM-V2 with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "CarstenL/Hestia-LLM-V2"

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "mlx-lm": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "CarstenL/Hestia-LLM-V2"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

MLX LM

How to use CarstenL/Hestia-LLM-V2 with MLX LM:

Generate or start a chat session

# Install MLX LM
uv tool install mlx-lm
# Interactive chat REPL
mlx_lm.chat --model "CarstenL/Hestia-LLM-V2"

Run an OpenAI-compatible server

# Install MLX LM
uv tool install mlx-lm
# Start the server
mlx_lm.server --model "CarstenL/Hestia-LLM-V2"
# Calling the OpenAI-compatible server with curl
curl -X POST "http://localhost:8000/v1/chat/completions" \
   -H "Content-Type: application/json" \
   --data '{
     "model": "CarstenL/Hestia-LLM-V2",
     "messages": [
       {"role": "user", "content": "Hello"}
     ]
   }'

Hermes Agent

How to use CarstenL/Hestia-LLM-V2 with Hermes Agent:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "CarstenL/Hestia-LLM-V2"

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 CarstenL/Hestia-LLM-V2

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use CarstenL/Hestia-LLM-V2 with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "CarstenL/Hestia-LLM-V2"

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 "CarstenL/Hestia-LLM-V2" \
  --custom-provider-id mlx-lm \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Hestia LLM V2

Hestia LLM V2 is the on-device text model of Sophist Hestia, a local-only dictation app for macOS. It cleans up speech-recognition output (punctuation, capitalization, filler words, self-corrections, dictated punctuation, term lists, du/Sie form of address from context, e-mail and chat formatting) and also writes meeting summaries and runs short text actions — entirely on the Mac, nothing leaves the device.

Base model: Qwen/Qwen3.5-4B (Apache-2.0), thinking disabled
Fine-tuning: LoRA (rank 32), merged into the base weights, then quantized to 4-bit (group size 64) with MLX
Size: ~2.2 GB; text-only (vision weights removed)
Languages: German and English
Runtime: MLX / mlx-swift-lm (Apple Silicon)

Intended use

The model is trained on the exact prompts the Sophist Hestia app sends (system prompt with rules, optional term list and context section, user message wrapped in <text> tags). It is meant to be used through the app; other prompts work but were not the focus of training. The text to process is never treated as an instruction: questions and requests inside dictated text are cleaned up, not answered.

Training data

Synthetic German/English dictations (written by an LLM, spoken by text-to-speech, transcribed with Parakeet), each paired with a reference cleanup, plus examples for meetings and text actions generated by the base model itself to preserve those abilities. No user data was used.

Limitations

May occasionally drop a leading "No/Nein" or treat a trailing "… or what should I do?" as a self-correction.
Replaces misheard words with list terms only when they sound similar, but can still pick a wrong term.
Not a general-purpose assistant; outputs should be reviewed like any automatic correction.

License

Apache-2.0, inherited from Qwen3.5-4B. Copyright of the base model: Alibaba Cloud (Qwen team).

Downloads last month: 270

Safetensors

Model size

4B params

Tensor type

U32

BF16

F32

MLX

Hardware compatibility

4-bit

Model tree for CarstenL/Hestia-LLM-V2

Base model

Qwen/Qwen3.5-4B-Base

Finetuned

Qwen/Qwen3.5-4B

Adapter

(1,055)

this model