Instructions to use hermitdave/Spark-X2.6-4B-MLX-SLQ-4.50bpw with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use hermitdave/Spark-X2.6-4B-MLX-SLQ-4.50bpw 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("hermitdave/Spark-X2.6-4B-MLX-SLQ-4.50bpw")

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 hermitdave/Spark-X2.6-4B-MLX-SLQ-4.50bpw with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "hermitdave/Spark-X2.6-4B-MLX-SLQ-4.50bpw"

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": "hermitdave/Spark-X2.6-4B-MLX-SLQ-4.50bpw"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

MLX LM

How to use hermitdave/Spark-X2.6-4B-MLX-SLQ-4.50bpw with MLX LM:

Generate or start a chat session

# Install MLX LM
uv tool install mlx-lm
# Interactive chat REPL
mlx_lm.chat --model "hermitdave/Spark-X2.6-4B-MLX-SLQ-4.50bpw"

Run an OpenAI-compatible server

# Install MLX LM
uv tool install mlx-lm
# Start the server
mlx_lm.server --model "hermitdave/Spark-X2.6-4B-MLX-SLQ-4.50bpw"
# Calling the OpenAI-compatible server with curl
curl -X POST "http://localhost:8000/v1/chat/completions" \
   -H "Content-Type: application/json" \
   --data '{
     "model": "hermitdave/Spark-X2.6-4B-MLX-SLQ-4.50bpw",
     "messages": [
       {"role": "user", "content": "Hello"}
     ]
   }'

Hermes Agent

How to use hermitdave/Spark-X2.6-4B-MLX-SLQ-4.50bpw 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 "hermitdave/Spark-X2.6-4B-MLX-SLQ-4.50bpw"

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 hermitdave/Spark-X2.6-4B-MLX-SLQ-4.50bpw

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use hermitdave/Spark-X2.6-4B-MLX-SLQ-4.50bpw with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "hermitdave/Spark-X2.6-4B-MLX-SLQ-4.50bpw"

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 "hermitdave/Spark-X2.6-4B-MLX-SLQ-4.50bpw" \
  --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"

Spark-X2.6-4B SLQ 4.50bpw

Statistically-lossless mixed-bitwidth quantization of XHToken/Spark-X2.6-4B using SLQ (COLM 2026).

Quantization summary

Average bitwidth: 4.4944 bpw (parameter-weighted)
Bitwidth distribution: 8-bit 16.7%, 5-bit 49.5%, 4-bit 33.8%
Artifact size: 2.821 GiB (BF16 base 7.657 GiB → 2.71× smaller)
Format: MLX packed checkpoint — integer grids stored as uint32, dequantized inside mx.quantized_matmul
Group size: 128
Calibration: FineWeb-Edu, 8,192 tokens, sequence length 512
Allocation: Per-layer greedy heuristic, target 4.5 bpw

How to run

mlx_lm.generate --model hermitdave/Spark-X2.6-4B-SLQ-4.50bpw --dtype float16

Or in Python:

from mlx_lm import load, generate
model, tokenizer = load("hermitdave/Spark-X2.6-4B-SLQ-4.50bpw", trust_remote_code=True)
response = generate(model, tokenizer, "Hello, how are you?", max_tokens=100)

trust_remote_code=True is required because mlx-lm imports the generated slq_model.py to build the architecture. Only load checkpoints you trust.

Validation

Perplexity (FineWeb-Edu, 512 sequences)

Model	PPL	Ratio
BF16	23.22	1.000
SLQ 4.50bpw	24.52	1.056

Zero-shot task recovery (100 examples/task, measured)

Task	BF16	SLQ 4.50bpw	Recovery
arc_challenge	0.4100	0.3600	0.8780
hellaswag	0.4000	0.4000	1.0000
winogrande	0.5500	0.5100	0.9273
mean			0.9351

Distribution metrics (FineWeb-Edu)

Metric	Value
Symmetric KLD	7.949
EAR	0.9129
KL divergence	0.0614

Citation

@inproceedings{helcig2026slq,
  title = {{Statistically-Lossless Quantization of Large Language Models}},
  author = {{Michael Helcig and Eldar Kurtic and Dan Alistarh}},
  booktitle = {{Conference on Language Modeling (COLM)}},
  year = {2026},
  url = {https://arxiv.org/abs/2605.02404},
}

Acknowledgments

Original SLQ implementation by IST-DASLab (Michael Helcig, Eldar Kurtic, Dan Alistarh)
Ported to MLX for Apple Silicon by Hermes Agent (Nous Research)
Quantized by hermitdave (HuggingFace: hermitdave)