compunerd/emerging-systems-models

Phox — BQSM Ternary Model Weights

Model weights for the Phox BQSM (Basin-Quotient-Machine) wave-rider inference engine.

Overview

This repository contains ternary-quantized model weights for the Phoenix Brain — a Kuramoto oscillator interference engine that performs inference through traveling wave activation rather than traditional matrix multiplication.

Architecture

• Engine: Phoenix Brain (phoenix_brain.c) — 4-ring Kuramoto core + demand-driven tendrils
• Weights: 4-level ternary quantization (80% sparsity, 2 bits per weight)
• Compute: AVX2 SIMD + OpenMP parallelism (CPU only)
• No GPU required: Engine is pure C with AVX2, runs on any x86-64 CPU

Models

File	Size	Description
gemma4-12b-ternary-normed.bqsm	2.8GB	Gemma-4-12B ternary weights (normalized)
gemma4-12b-bf16-ternary.bqsm	1.2GB	Gemma-4-12B BF16-derived ternary
gemma4-12b.vocab	2.8MB	Token vocabulary
hermes-3b-ternary.bqsm	766MB	Hermes-3-3B ternary weights
hermes-3b.vocab	1.3MB	Token vocabulary
phoenix-evolved.pbrain	~2MB	Evolved brain state file

Files

• gemma4-12b-ternary-normed.bqsm — Main 12B parameter model with normalized weights
• tokenizer/ — Gemma 4 tokenizer files (compatible with HuggingFace transformers)
• gemma4-12b.vocab — Token-to-ID vocabulary mapping for the C engine

Usage

Compile the engine:

cc -O3 -std=c11 -march=native -fopenmp phoenix_brain.c -o phoenix -lm

Run inference:

./phoenix gemma4-12b-ternary-normed.bqsm 20

Chat mode (with ring buffer I/O):

./phoenix gemma4-12b-ternary-normed.bqsm --chat

Performance

• ~4.9 tok/s on 12B model (12-core CPU, AVX2)
• Model loaded via mmap (no RAM duplication)
• 48 layers of wave-rider physics per token

Citation

If you use this work, please reference:

• Nicholas Bumgarner, emerging.systems
• "Basin-Quotient-Machine: Cross-Harmonic Computation"