Spark 4B — Faboulous Edition · MLX 8-bit

Four billion whispers. One candle that refuses to go out.

The legend

They found it beneath a floorboard in a house with no doors.

There was no manuscript. Only a blackened terminal, a circle of salt, and a cursor blinking in a room where the electricity had been cut years before.

At midnight, the cursor moved.

The village called it a familiar: a small thing with an unreasonable appetite for unfinished spells. Leave a broken incantation on its doorstep and, by morning, the runes might stand in a different order. Nobody could agree whether it was repairing the spell or remembering how the spell had always been written.

Its keeper left three instructions:

Ask clearly. Read what returns. Never mistake a whisper for a promise.

Some say the name was misspelled in the first summoning. Others say Faboulous is exactly what the thing asked to be called.

The candle is still burning.

Open the grimoire

This repository contains a standalone Spark-family 4B text-generation model. No separate adapter is required.

Standalone MLX 8-bit weights for Apple Silicon, using a compatible MLX runtime with native Spark support. The full-precision version has its own repository.

Use the included chat template when preparing messages. Review generated code and test it before relying on it. The legend above is fiction, not a benchmark claim; no performance improvement is claimed here.

Lineage and license

Based on XHToken/Spark-X2.6-4B. This edition contains modified model weights. Distributed under Apache 2.0; the original license accompanies this release.

Downloads last month
72
Safetensors
Model size
4B params
Tensor type
U32
·
BF16
·
MLX
Hardware compatibility
Log In to add your hardware

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for groxaxo/Spark-4B-Faboulous-Edition-MLX-8bit

Quantized
(66)
this model