🤫 Puppy One on Apple silicon
Personal AI should feel personal: responsive, quiet, and close. Puppy One pairs the new generation of Apple silicon with 🤫 Private Agent One - guided day-one activation included, consent-led, and tuned for the work you actually do.
The 2026 collection
Start with the work. Choose the smallest Puppy that can keep the whole job local. That is where price, performance, privacy, and energy meet.
Configured on Mac mini · M6
Everyday private agent, retrieval, voice, and 8B-14B local models.
The smallest complete machine. Our expected token-per-watt leader for personal, always-on work - pending qualification.
Configured on Mac mini · M5 Pro
Advisors, builders, and 32B-class local models with useful context.
More memory and bandwidth without leaving the five-inch footprint. A quiet, always-on practice agent.
Configured on Mac Studio · M5 Max
70B-class local models, coding agents, richer context, and concurrency.
The strongest balance of model headroom, bandwidth, power, and capital cost for demanding personal AI.
Configured on Mac Studio · M5 Ultra
Very large open-weight models, research, multi-agent teams, and clusters.
Buy it for memory fit and absolute local capability - not because the biggest chip automatically means the best efficiency.
A better efficiency standard
Agentic systems read, reason, retrieve, call tools, wait, and act. A single tokens-per-watt number hides that journey. We will publish the number - and the outcome behind it.
Prefill and decode measured separately at the wall, not inferred from chip marketing.
Successful, human-accepted agent jobs per unit of energy - our primary outcome metric.
The share of useful work completed without sending private context away from the device.
Prompt processing, first token, final answer, and tool latency measured as one experience.
🤫 research note · August 2026
Apple's new systems change the local-AI frontier in two different ways. M6 is the efficiency story: a 2-nanometer design, dual Neural Engines, GPU Neural Accelerators, and up to 170GB/s in a five-inch desktop. M5 Ultra is the capacity story: up to 512GB of unified memory and 1.2TB/s of bandwidth, enough to keep models with hundreds of billions of parameters on one desk.
Between them, M5 Pro and M5 Max are where most professional customers should live. The 64GB mini is the compact 32B machine. The 128GB Studio is our high-confidence recommendation for 70B-class models and concurrent agents. Ultra earns its place only when the workload needs its memory.
Read the full research noteMeasured, modeled, or unknown
Apple reports relative "LLM prompt processing" gains, but its product pages define that test as time to first token. Apple does not publish the tested model, quantization, context length, sustained decode rate, or workload wall power. Its technical specifications publish maximum continuous system power - 155W for Mac mini and 480W for Mac Studio - not inference power. Any exact Puppy token-per-watt figure today would be false precision.
Our launch qualification will record wall energy, time-to-first-token, prompt and decode throughput, thermals, quality, tool success, and idle draw for every recommended model/configuration pair. Until that matrix passes, efficiency labels on this page are recommendations, not benchmark claims.
Private intelligence, ready when you are.
Reserve any Puppy for any amount from $0.01. Your reservation is refundable and credited toward purchase.
And a straight deal on price: before you pay, hussh checks the lowest advertised U.S. price for your exact Apple configuration - Apple, Costco, and the major retailers - and prices your hardware below it.2
Reserve Puppy One