Syzygy
Syzygy builds Mach Studio, a local-first desktop coworking application for Apple Silicon Macs that runs open models entirely on-device with no cloud dependency. The company's Mach Engine uses an additive compression scheme that retains 96.3% of each model's BF16 teacher performance across twelve benchmarks, fitting 125.7B text parameters in 26.7 GB. The product supports agentic workflows across Slack, Gmail, and Notion, with a TypeScript SDK providing deterministic permission policies and durable checkpoints for reliable long-running tasks.
- Artificial Intelligence
- AI Agents
- Developer Tools
- Software Only
Funding
Founders
Founder details are not available yet.
Product
Problem
Running capable AI models locally on personal hardware typically requires significant memory and compute resources, forcing users to either accept heavily quantized models with degraded quality or rely on cloud infrastructure that introduces privacy risks, latency, and recurring costs. The emerging class of local agentic workflows—where models read, draft, and act across applications—lacks a runtime designed for the constraints of small models and finite context windows, leading to unreliable long-running tasks and inefficient tool usage.
Solution
Syzygy provides Mach Studio, a desktop coworking application for Apple Silicon Macs that runs open models entirely on-device with no accounts, API keys, or cloud dependency. The companion Mach Engine delivers a mixture-of-experts inference engine using an additive compression scheme that retains 96.3% of each model's BF16 teacher performance across twelve benchmarks. The product lets agents read, draft, and act across connected apps like Slack, Gmail, and Notion, drive the browser through computer-use actions, and work alongside users in shared documents. A built-in agents SDK with deterministic permission policies, durable checkpoints, and context-compaction techniques keeps small local models reliable during long-running tasks.
Target Audience
Primary users are individual developers, knowledge workers, and small teams on Apple Silicon Macs who want private, offline access to capable local models for agentic coding, personal automation, and document work without cloud data exposure.
Features
- Mach 2 Additive Medium compresses Qwen3.8-Flash-Next to 1.70 bits per weight, fitting 125.7B text parameters in 26.7 GB while retaining 96% of BF16 benchmark scores on a single 32 GB GPU
- Add/subtract-only integer-code weight matmuls with trained scale surfaces distilled via on-policy distillation, eliminating multiplication operations during inference
- Mach Agents TypeScript SDK with deterministic permission policies that match inside tool arguments (headers, paths, shell calls) to allow, deny, or pause-for-approval before execution, plus hard ceilings on tokens, dollars, and wall-clock time
- Context management featuring per-turn toolset search (avoiding full catalogs), compacted history, size-capped working memory, isolated sub-agent contexts, and relevance filtering
- Automatic JSON repair loops and structured-output validation with retry rounds that tolerate malformed model output without crashing
- Native integration with Claude Code, Codex, OpenCode, and Hermes Agent, plus scheduled recurring tasks and delegation mode for bounded sub-agent work