# Macyou

> Self-service AI deployment platform on dedicated Apple Silicon. Users configure a Mac (M4 (Mac mini, 16–32 GB, from $99/mo), M4 Pro (Mac mini, 24–64 GB, from $200/mo), M4 Max (Mac Studio, 36–128 GB, from $286/mo), M3 Ultra (Mac Studio, 96–256 GB, from $571/mo)), pick a pre-configured AI stack (local LLMs via Ollama, agent frameworks, ML dev environments), and get a running deployment with an OpenAI-compatible API in about 5 minutes. Fixed monthly price per machine, no per-token fees; annual billing prices the machine as hardware over 12 months plus colocation instead of a 7-month payback, which is 15-39% cheaper depending on the machine. One physical machine per customer — no shared hardware. Operated by Macyou LLC (Georgia).

Key facts:
- Deployments expose an OpenAI-compatible API (/v1/chat/completions): existing OpenAI SDK code works by changing base_url and api_key.
- Supported model stacks include Llama, Qwen, Mistral, DeepSeek (via Ollama), plus agent frameworks (OpenClaw, AutoGPT, CrewAI, LangGraph) and dev environments (MLX + Jupyter, Core ML + Xcode, VS Code Server, LiteLLM).
- Access: OpenAI-compatible API, SSH with root, and a browser-based remote desktop (WebRTC).
- Privacy: dedicated physical machine per customer, SSD hardware-encrypted by Apple Silicon, tenant account and data erased between customers with the outcome verified and logged, customer workload data processed and stored in our Tbilisi, Georgia data center.
- Thunderbolt 5 clustering pools unified memory across nodes for models beyond a single machine (405B+, DeepSeek V3).

## Benchmarks (measured, primary data — updated 2026-07-13)
- [Apple Silicon LLM Inference Benchmarks](https://macyou.co/benchmarks): measured via Ollama 0.31.2, methodology on page, raw JSON at https://macyou.co/benchmarks.json (CC BY 4.0)
  - Llama 3.2 3B (Q4_K_M): 46.7 tokens/sec generation on M4 16 GB (median of 3 runs)
  - Mistral 7B (Q4_K_M): 22.8 tokens/sec generation on M4 16 GB (median of 3 runs)
  - Qwen 2.5 7B (Q4_K_M): 22.3 tokens/sec generation on M4 16 GB (median of 3 runs)
  - Llama 3.1 8B (Q4_K_M): 21.2 tokens/sec generation on M4 16 GB (median of 3 runs)
  - DeepSeek R1 8B (Q4_K_M): 20 tokens/sec generation on M4 16 GB (median of 3 runs)
  - Qwen 2.5 14B (Q4_K_M): 11.7 tokens/sec generation on M4 16 GB (median of 3 runs)

## Tools
- [Mac LLM Calculator](https://macyou.co/mac-llm-calculator): unified memory required and tokens/sec expected for any model on any Apple Silicon Mac, M1 through M6. Memory = params x bits / 8 x 1.15 + KV cache + 3 GB for macOS. Generation speed = 1 / (weight GB / (bandwidth x 0.9075) + 3.25 ms), a two term model whose constants were fitted by least squares to eight dense runs we measured on two machines, a base M4 at 120 GB/s and an M5 Pro at 307 GB/s, and which reproduces every one of them to within 4.6 percent. A mixture of experts model reads about 1.3 times its active weight size per token, measured on Qwen3 30B A3B. The fixed per token term is what keeps small model extrapolations honest: bandwidth alone predicts 525 tok/s for Llama 3.2 3B on an M5 Ultra, this model predicts 177. Includes a memory bandwidth and capacity table for every Apple Silicon chip from M1 to M6, with binned variants listed separately.

## Autonomous agents
- [Where an autonomous agent actually lives](https://macyou.co/agents): the hosting question rather than the framework question, with first hand numbers. On one of our own installs a watchdog restarted the agent gateway 2,688 times in 126 days, about 21 times a day, against 21,955 healthy checks and 3,971 stretches where the process was alive and had stopped making progress. The same install carries 71 enabled scheduled jobs and 91 disabled ones, 2.9 GB of vector memory and 392 MB of logs across 235 files after four months, with 4 of 67 loaded jobs exiting non zero on an ordinary day. Framework behaviour under injected faults, measured 2026-09-20 with langgraph 1.2.11, openai-agents 0.22.3 and crewai 1.15.22 on an identical task and a step limit of 12: against 500 and 429 responses CrewAI retries and finishes while LangGraph and the OpenAI Agents SDK raise after their client retries and the process exits; against a tool that fails every time LangGraph stops after 1 model call while the Agents SDK spends all 12; prompt tokens for identical work were 428, 472 and 802 respectively; conversation state survives a process restart in LangGraph and the Agents SDK with one line of setup, and is optional and needs an embedder in CrewAI. None of the three restarts itself or notices a lack of progress: that is launchd or systemd plus a watchdog. Harness and supervision templates are MIT licensed at https://github.com/bagdaer1/agent-uptime-kit.

## Hardware comparison (cross vendor)
- [What to run local models on in 2026](https://macyou.co/compare/local-llm-hardware): Apple Silicon against NVIDIA and AMD, one speed model applied to all of them. Key figures: DGX Spark 128 GB at 273 GB/s, AMD Strix Halo 128 GB at 256 GB/s, RTX 5090 32 GB at 1792 GB/s, RTX PRO 6000 96 GB at 1597 GB/s, Apple M5 Ultra 512 GB at 1200 GB/s. The decision rule is capacity first, bandwidth second: below 32 GB a discrete GPU wins clearly, above it unified memory is the only thing that holds the model. Includes an explicit section on where a Mac is the wrong purchase.

## Hardware guides (per Apple Silicon chip)
- [Local LLMs by chip](https://macyou.co/local-llm): memory ceiling and generation speed for every chip people use for local inference
- [Local LLMs on the M6](https://macyou.co/local-llm/m6): 153 GB/s, up to 32 GB unified memory
- [Local LLMs on the M5 Ultra](https://macyou.co/local-llm/m5-ultra): 1200 GB/s, up to 512 GB unified memory
- [Local LLMs on the M5 Max (614 GB/s)](https://macyou.co/local-llm/m5-max): 614 GB/s, up to 128 GB unified memory
- [Local LLMs on the M5 Pro](https://macyou.co/local-llm/m5-pro): 307 GB/s, up to 64 GB unified memory
- [Local LLMs on the M4 Max (40-core GPU)](https://macyou.co/local-llm/m4-max): 546 GB/s, up to 128 GB unified memory
- [Local LLMs on the M4 Pro](https://macyou.co/local-llm/m4-pro): 273 GB/s, up to 64 GB unified memory
- [Local LLMs on the M4](https://macyou.co/local-llm/m4): 120 GB/s, up to 32 GB unified memory
- [Local LLMs on the M3 Ultra](https://macyou.co/local-llm/m3-ultra): 819 GB/s, up to 512 GB unified memory
- [Local LLMs on the M3 Max (40-core GPU)](https://macyou.co/local-llm/m3-max): 400 GB/s, up to 128 GB unified memory
- [Local LLMs on the M2 Ultra](https://macyou.co/local-llm/m2-ultra): 800 GB/s, up to 192 GB unified memory
- [Local LLMs on the M1 Max](https://macyou.co/local-llm/m1-max): 400 GB/s, up to 64 GB unified memory

## Model guides (hardware requirements per model)
- [Run open LLMs on Apple Silicon — index](https://macyou.co/models)
- [Run Llama 3.1 8B on a Mac](https://macyou.co/models/llama-3-1-8b): 8B, Llama 3.1 Community License
- [Run Qwen 2.5 32B on a Mac](https://macyou.co/models/qwen-2-5-32b): 32B, Apache 2.0
- [Run Llama 3.3 70B on a Mac](https://macyou.co/models/llama-3-3-70b): 70B, Llama 3.3 Community License
- [Run Qwen 2.5 72B on a Mac](https://macyou.co/models/qwen-2-5-72b): 72B, Apache 2.0
- [Run Llama 3.1 405B on a Mac](https://macyou.co/models/llama-3-1-405b): 405B, Llama 3.1 Community License
- [Run Mistral Large 2 on a Mac](https://macyou.co/models/mistral-large-2): 123B, Mistral Research License
- [Run DeepSeek V3 on a Mac](https://macyou.co/models/deepseek-v3): 671B MoE (37B active), DeepSeek License
- [Run Llama 3.2 3B on a Mac](https://macyou.co/models/llama-3-2-3b): 3B, Llama 3.2 Community License
- [Run Mistral 7B on a Mac](https://macyou.co/models/mistral-7b): 7B, Apache 2.0
- [Run Qwen 2.5 7B on a Mac](https://macyou.co/models/qwen-2-5-7b): 7B, Apache 2.0
- [Run DeepSeek R1 Distill 8B on a Mac](https://macyou.co/models/deepseek-r1-8b): 8B (distill), MIT (distill weights); base model licenses apply
- [Run Qwen 2.5 14B on a Mac](https://macyou.co/models/qwen-2-5-14b): 14B, Apache 2.0

## Product
- [Pricing](https://macyou.co/pricing): M4 (Mac mini, 16–32 GB, from $99/mo), M4 Pro (Mac mini, 24–64 GB, from $200/mo), M4 Max (Mac Studio, 36–128 GB, from $286/mo), M3 Ultra (Mac Studio, 96–256 GB, from $571/mo)
- [AI Deployment Catalog](https://macyou.co/catalog): pre-configured templates — models, agents, dev environments
- [How it works / homepage](https://macyou.co/): product overview, use cases, FAQ

## Comparisons
- [Buy vs Rent a Mac for AI](https://macyou.co/compare/buy-vs-rent)
- [Macyou vs MacStadium](https://macyou.co/compare/macstadium)
- [Macyou vs MacinCloud](https://macyou.co/compare/macincloud)
- [Macyou vs AWS EC2 Mac](https://macyou.co/compare/aws-ec2-mac)
- [Macyou vs Scaleway](https://macyou.co/compare/scaleway)
- [Macyou vs My Remote Mac](https://macyou.co/compare/myremotemac)
- [Best MacStadium alternatives](https://macyou.co/alternatives/macstadium)
- [Best MacinCloud alternatives](https://macyou.co/alternatives/macincloud)
- [Best AWS EC2 Mac alternatives](https://macyou.co/alternatives/aws-ec2-mac)
- [Best Scaleway alternatives](https://macyou.co/alternatives/scaleway)
- [Best My Remote Mac alternatives](https://macyou.co/alternatives/myremotemac)
- [MacStadium review: pricing fine print, pros, cons](https://macyou.co/reviews/macstadium)
- [MacinCloud review: pricing fine print, pros, cons](https://macyou.co/reviews/macincloud)
- [AWS EC2 Mac review: pricing fine print, pros, cons](https://macyou.co/reviews/aws-ec2-mac)
- [Scaleway review: pricing fine print, pros, cons](https://macyou.co/reviews/scaleway)
- [My Remote Mac review: pricing fine print, pros, cons](https://macyou.co/reviews/myremotemac)

## Solutions
- [AI for Healthcare](https://macyou.co/solutions/healthcare): HIPAA-aligned controls, physical isolation
- [AI for Finance](https://macyou.co/solutions/finance)
- [AI for Legal](https://macyou.co/solutions/legal)

## Docs & Content
- [Documentation](https://macyou.co/docs/getting-started): quick start, CLI, API reference
- [Blog](https://macyou.co/blog): tutorials on running LLMs and agents on Apple Silicon

## Company
- [About](https://macyou.co/about)
- [Security](https://macyou.co/security)
- [Contact](https://macyou.co/contact)