---
language:
- en
- ar
license: mit
tags:
- silx-ai
- quasar
- foundation-model
- 3b
- moe
- long-context
- bittensor
- sn24
- distillation
- hybrid-transformer
pipeline_tag: text-generation
library_name: transformers
---

<p align="center">
  <img src="./Quasar.png" alt="Quasar Foundation Model" width="100%">
</p>

# **Quasar Foundation Models (RoPE Base)**

**Quasar Foundation Models** are SILX AI’s core models designed for **long-context reasoning**, **agentic systems**, and **persistent memory-based intelligence**.

This release is **NOT a state-of-the-art final model**.  
It is a **base pretraining model** designed specifically for **distributed knowledge distillation on Bittensor (SN24 Quasar subnet)**.

The goal is to create a shared architecture where miners continuously **distill knowledge from frontier models (e.g., Qwen, GLM)** into Quasar.

---

## ⚠️ Important Note

This model is:

- A **base model**
- **Pretrained for only a few billion tokens**
- Designed for **distillation and scaling**, not benchmarking

Performance will improve through **iterative subnet training + distillation cycles**.

---

## Model Overview

- **Model Name:** Quasar 3B (RoPE Base)
- **Organization:** SILX AI
- **Architecture:** Quasar-RoPE Hybrid Transformer
- **Total Parameters:** 3B  
- **Active Parameters:** ~1B (Mixture-of-Experts)
- **Training Stage:** Stage 1 (Base Pretraining)
- **Sequence Length:** 16K tokens (RoPE phase)

---

## Training Strategy

Quasar follows a **multi-stage training pipeline**:

### **Stage 1 — RoPE Pretraining**
- Train using **Rotary Positional Embeddings (RoPE)**
- Context length: **16K tokens**
- Objective: stabilize training and build core reasoning

### **Stage 2 — Distillation (SN24)**
- Distributed training on **Bittensor subnet (SN24)**
- Miners distill knowledge from:
  - Qwen
  - GLM
- Target: transfer reasoning + capabilities into Quasar

### **Stage 3 — DroPE Long-Context Training**
- Remove positional embeddings entirely (**DroPE phase**)
- Transition to **position-free reasoning**
- Train on **ultra-long context (up to 5M tokens)**

This staged approach allows:
- Stable early training
- Efficient knowledge transfer
- Extreme context scaling without positional bottlenecks

---

# **Quasar-RoPE Hybrid Architecture**

Quasar is a **high-throughput hybrid transformer** designed for **trillion-token scale training**.

It combines:
- **Looped computation**
- **Persistent latent memory**
- **Hybrid attention mechanisms**
- **Stable Mixture-of-Experts routing**

---

## 1. Looped Transformer Logic

Instead of increasing depth traditionally, Quasar uses **looped execution**:

- A fixed set of layers is reused multiple times (`num_loops`)
- This multiplies effective depth without increasing VRAM

### Key Mechanism:

- **Anchor P (Input Injection):**
  - Embedding output is stored as `P`
  - Injected into the hidden state at every loop
- **Gradient Stabilization:**
  - Injection gradients scaled by `1 / num_loops`
  - Prevents instability during recirculation

---

## 2. Hybrid Layer Composition

Each loop contains a mix of:

### **Quasar Layers**
- Use **Latent Memory Module**
- Handle long-range dependencies
- Read/write persistent state

### **GLA Layers (Gated Linear Attention)**
- Fast, RNN-like recurrence
- Efficient local sequence modeling

---

## 3. Persistent Latent Memory

A defining component of Quasar:

- **Memory Slots:**
  - Fixed parameter banks (e.g., 128–256 slots)

- **Segment Compression:**
  - Tokens grouped into segments (default: 64 tokens)
  - Reduced noise during updates

- **Saliency Gating:**
  - Learns which information is important
  - Writes only high-value signals to memory

---

## 4. SMEBU (Stability-Maximized Expert Balancing Unit)

Custom Mixture-of-Experts system:

- **Global Bias Buffers**
  - Stored outside optimizer
  - Prevent routing collapse

- **Zero-Loop Updates**
  - Expert balancing done in vectorized pass
  - No recursive instability

- **Sparse Activation**
  - ~1B active parameters per forward pass

---

## 5. Technical Specifications

- **Normalization:** RMSNorm (Pre-Norm)
- **Positional Encoding:** RoPE (`theta = 1,000,000`)
- **Initialization:** Depth-scaled `1/sqrt(2L)`
- **Architecture Type:** Hybrid Transformer + Memory + MoE

---
# Architecture Overview

## Core Data Flow

```
Token IDs
  ↓
Embedding Layer
  ↓
Anchor P Snapshot
  ↓
┌──────────────────────────────────────────────┐
│ Loop (i < num_loops)                         │
│                                              │
│   Quasar Block                               │
│        ↓                                     │
│   GLA Block                                  │
│        ↓                                     │
│   SMEBU MoE                                  │
│        ↓                                     │
│   Inject Anchor P (Residual Conditioning)    │
└──────────────────────────────────────────────┘
  ↓
Next Loop Iteration (state updated)

Final Loop Output
  ↓
RMSNorm
  ↓
LM Head
  ↓
Logits
```

---

## Latent Memory Update Path

```
Hidden States
  ↓
Layer Normalization (RMSNorm)
  ↓
Segment Compressor
  ↓
Segment Representation (Z)
  ↓
  ├──────────────→ Saliency Gate (importance scoring)
  │                        ↓
  │                     Write Signal
  │                        ↓
  └──────────────→ Memory Write Operation
                           ↓
              Persistent Memory Bank (M)
                           ↓
                  Updated Memory (M')
                           ↓
                  Memory Read Module
                           ↓
              Memory-Augmented Hidden State
                           ↓
                         Output
```

---

## SMEBU MoE Stability Flow

```
Router Network
  ↓
Token Routing Scores
  ↓
  * Global Bias Buffer (non-trainable stability path)
  ↓
Top-K Expert Selection
  ↓
Selected Experts
  ↓
Expert Output Aggregation
  ↓
Final MoE Output
  ↓
Post-Loop Bias Update (vectorized, stabilized)
```

---

# Intended Use

This model is designed as a **foundation base model** for the Quasar ecosystem and is primarily intended for:

- **Bittensor SN24 miners** participating in distributed training and knowledge distillation
- **Distillation pipelines** transferring capabilities from frontier models (e.g., Qwen, GLM)
- **Research on long-context architectures**, especially beyond traditional positional encoding limits
- **Agentic system development**, where persistent memory and long-horizon reasoning are required

---

# Next Steps

- Training on **SN24** in the coming days
- Miners distill knowledge into this model
- Then we go for **Run 2 — DroPE training** at **5M tokens**