monkiey/StarSupernova

Star Supernova: Gemma-4 MoE (12 Experts, Sparse Top-4 Routing, 4-bit)

Star Supernova is a high-capacity, sparse Mixture-of-Experts (MoE) model built on Gemma-4 with 12 routed experts and a shared dense feed-forward network. It dynamically activates Top-4 experts per token, expanding total model capacity to ~13.47 Billion parameters while maintaining high inference speed and zero swap thrashing on 16 GB Apple Silicon Macs.

Quantized to 4-bit affine precision, it runs natively on Apple Silicon Metal GPU via MLX with a memory footprint of only ~8.7 GB VRAM, leaving >3.4 GB of free GPU headroom under macOS's 12.1 GB Metal limit.

---

Architecture & Specifications

Feature	Specification
Model Name	Star Supernova
Base Architecture	Gemma-4 (35 Layers, Hidden Dim 1536)
Total Parameters	~13.47 Billion
Active Parameters / Token	~5.54 Billion (41.1% active / 58.9% sparse)
Routing Mechanism	Sparse Top-4 Routing across 12 Experts + 1 Shared Dense MLP
Quantization	4-bit affine quantization (group_size=64, routers in 8-bit)
Memory Footprint	~8.7 GB VRAM (Leaves >3.4 GB free on 16 GB Macs)
Inference Framework	MLX / Unsloth Studio / Apple Silicon Metal GPU
Context Length	131,072 tokens

---

Key Features

1. Massive Capacity, Compact Footprint:

   • Total expert parameter capacity expanded from 4 to 12 experts.
   • Operates in ~8.7 GB of Metal VRAM, allowing the entire model and KV cache to run purely in physical RAM without relying on macOS SSD virtual swap.

2. Dynamic Top-4 Sparse Routing:

   • Each token routes through the top 4 most specialized experts in each layer in addition to the shared base MLP.
   • High parameter density and expressive specialization across math, logic, creative writing, and systems engineering.

3. Native CoT Reasoning:

   • Native <|channel>thought Chain-of-Thought reasoning support.

---

Quickstart with MLX

Installation

pip install mlx-lm

Generation

from mlx_lm import load, generate

model, tokenizer = load("monkiey/StarSupernova")

prompt = "<start_of_turn>user\nExplain how sparse Mixture-of-Experts routing optimizes memory and throughput.<end_of_turn>\n<start_of_turn>model\n"

response = generate(
    model,
    tokenizer,
    prompt=prompt,
    max_tokens=512,
    verbose=True
)
print(response)

---

Deployment in Unsloth Studio

To use Star Supernova in Unsloth Studio:

1. Clone or download this repository into your Unsloth Studio outputs directory:

   git clone https://huggingface.co/monkiey/StarSupernova ~/.unsloth/studio/outputs/Star_Supernova
   

2. Open Unsloth Studio and select Star Supernova from the model dropdown.
3. Recommended inference settings:
   • Temperature: 0.3
   • Top-P: 0.9
   • Repetition Penalty: 1.15
   • Max Tokens: 17280

---

License

Apache 2.0