QuilaQualiaTrace Language Model

Novel dynamic graph reasoning system with 161B effective parameters

Get Started

View on GitHub

Dynamic Graph Architecture

32,768 neurons with dynamic topology evolution, ~19% sparse activation

6-Phase Inference Pipeline

Replay, LastInput, Context Re-read, Plan, Think, Output with Higher-Order Thinking

Multimodal Support

Text, image, audio, video via VQ-GAN encoding boundary

Multi-GPU Distributed

CUDA Unified Memory with NCCL for efficient multi-GPU training/inference

Quick Start

bash

# Clone repository
git clone https://github.com/rand0mdevel0per/sxlm.git
cd sxlm

# Build
cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release

# Run API server
python src/python/run_server.py

Architecture Highlights

161B effective parameters (13B explicit + 148B implicit KFE)
Hybrid attention system with NSA, SSM, Linear Attention, DRC
5-tier memory hierarchy (GPU SRAM → NVMe disk)
Consistency Model for 1-step denoising