# Syzygy: local models and agent work on Apple Silicon

Canonical: https://mudpie.ai/companies/syzygy/
Breadcrumb: [Home](https://mudpie.ai/) / [Companies](https://mudpie.ai/companies/) / [Syzygy: local models and agent work on Apple Silicon](https://mudpie.ai/companies/syzygy/)
Author: Ali Abouelatta (https://mudpie.ai/authors/ali-abouelatta/)
Published: 2026-09-20
Updated: 2026-09-20
Research type: Company profile
Method: Public-source research. Official Speedrun and company pages checked September 20, 2026. Product claims are attributed to their sources.

Syzygy describes local-model tooling that connects apps and browsers to developer agents while reducing dependence on cloud inference.

## What it does

The company homepage describes Mach Studio as a local-model desktop app for Apple Silicon that connects apps and browsers and supports Claude Code, Codex, OpenCode and Hermes Agent. The benchmark page reports a company-published Mach-1 Small result of 96.3% mean score retention across twelve benchmarks.

## Buyer and task

Developers and teams that need local inference, lower cloud dependence or agent work across their Mac apps.

## Workflow boundaries

The benchmark is a published result, not an independent reproduction. The roster website maniac.ai redirects to the public withsyzygy.com surface; that redirect is recorded rather than treated as a second identity.

## What I would ask

Ask which models and Apple chips are supported, what leaves the machine, and how local agents are isolated from the user’s files.

## Why it fits

Syzygy has two connected surfaces in the checked pages. Mach Studio is a free beta desktop app for Apple Silicon that runs open models on-device, connects to apps and browsers, schedules recurring work and exposes the same local models to tools such as Claude Code, Codex, OpenCode and Hermes Agent. The Mach engine and benchmark pages explain the model-compression side: Mach-1 Small is described as a 35B-A3B model at about 1.7 bits per weight, with a published mean retention of 96.3% across twelve benchmark-level ratios against its own BF16 teacher.

That is a coherent choice for a developer who cares about local execution, model control or reducing cloud dependence. The benchmark number needs to stay in its proper category. Retention is a ratio against a teacher, not an absolute score or an independent guarantee of useful local work. The source page also lists local wall-clock comparisons and identifies the comparison arms and methodology; it does not replace a buyer’s own workload test.

The decision gate is operational: confirm the supported Apple Silicon machines, model library, memory requirements, what data leaves the Mac and how an agent is limited when it can read or act across local apps. The roster’s `maniac.ai` address redirects to the Syzygy pages used here; this profile preserves that identity trail.

## Quick facts

| Field | Sourced detail |
| --- | --- |
| Buyer | Developers and teams running local agent workflows on Apple Silicon |
| Product | Mach Studio desktop app plus Mach inference engine |
| Public access | Free Apple Silicon beta; no account, API key or cloud named on the product page |
| Published result | 96.3% mean retention across twelve benchmarks, a vendor-published comparison |

## Sources checked

[official Speedrun profile](https://speedrun.a16z.com/companies/syzygy) · [company homepage](https://withsyzygy.com/) · [source page](https://withsyzygy.com/docs/mach/benchmarks) · [source page](https://withsyzygy.com/product)

Sources checked — September 20, 2026.


## Author disclosure

I cofound Lazyweb and publish Mudpie. This is an owner-written publication, not an independent testing organization. Research notes distinguish observations, sourced reporting and editorial judgment.