S-MoE

⚡

Model-Agnostic

Run any fine-grained Mixture of Experts model — Qwen3-235B, DeepSeek, and beyond. S-MoE auto-detects architecture topology at boot time. No recompilation, no hardcoded shapes.

🧲

Phonon Streaming

Inspired by Filippo Biondi's SAR acoustic imaging. Routing is computed exactly from the model's own hidden state; the retained expert ring and each layer's ranked gate echo keep the Streamer's pread() calls ahead of the GPU's appetite.

🔒

Zero-Allocation Loop

No malloc. No new. No std::vector::resize. Inside the token generation loop, memory is a river — pre-carved at startup, flowing without interruption. Every buffer is pre-allocated and 16KB Metal-aligned.

🍎

Apple Silicon Native

Direct I/O via F_NOCACHE bypasses the OS page cache entirely. Metal zero-copy UMA bridging means expert weights arrive from NVMe and land directly onto GPU registers. The hardware is used as intended.

🌍

Democratic by Design

A 32 GB MacBook runs the same 235B frontier intelligence as a 512 GB Mac Pro — just slower. Speed scales with hardware. Quality does not degrade. This is AI access as a right, not a luxury.

🔬

Open Architecture

Every byte, every struct field, every atomic flag is documented. S-MoE is built to be understood, forked, extended, and challenged. The Monolithic Delusion ends here.

What is this?

S-MoE is an experiment, not a product.

It was built by a self-taught web developer — Luca Visciola (@melasistema) — whose GitHub is full of APIs and web stacks, not GPU kernels and low-level memory management. The architecture was designed in collaboration with an AI agent. I brought the curiosity, the "what if", and the direction; the agent brought the C++ implementation depth.

The core idea is a philosophical translation of the work of Ing. Filippo Biondi, an Italian engineer who in 2022 published a stroke of genius applying Synthetic Aperture Radar (SAR) to map the interior of the Great Pyramid of Giza. His key insight: Electromagnetic waves cannot penetrate solid rock, but when the EM pulse strikes the surface, it generates acoustic phonons — mechanical sound vibrations that propagate through the stone. Biondi used a brilliant Doppler technique and Fourier transforms to measure the resulting sub-nanometer surface micro-movements caused by the returning phonons. From those surface vibrations alone, he generated a Tomographic image, capturing all the 3D information of the pyramid's interior projected onto a 2D plane. You cannot see inside the rock, so you measure the sound the rock makes when you strike it.

In S-MoE: the NVMe SSD is the rock. The Surface Scout is the surface strike. The predicted expert activations are the phonons.

The claim is simple: if only ~5% of a frontier model fires per token, you should only need ~5% of the model in RAM at any time. Speed will scale with hardware. Quality should not degrade.

Whether that claim holds at 235B parameter scale is what we are trying to find out.

This is not a research paper. There are no guarantees. There are bugs. The code is public, the philosophy is public, and the mistakes are public. If you find a better way, please make a PR.

S-MoESeismic Mixture of Experts