Vision
Our long-term ambition is personal AGI and, eventually, superintelligence.
We want each person to have a private intelligence that understands them, works on their behalf and can ultimately live on hardware they control. That ambition extends across the agent, its digital brain, our own advanced models and the hardware that runs them.
Where we are
Being far beyond today’s capabilities does not make something impossible.
The north star is broadly capable personal intelligence that can understand its owner’s world, reason across domains and carry out increasingly complex work under the owner’s control. The longer ambition is personal access to superintelligence, with privacy and ownership carried through to the full system.
The Axel pod
A dedicated home for Axel.
A frontier-class intelligence and a lifetime of memory in one machine, running on silicon designed for the way it thinks.
A local pod that runs the model, agent and digital brain, with the owner’s phone connecting remotely. Over time, we want people to control the intelligence they rely on and the place where their personal context lives. What we learn guides custom hardware and, eventually, chip design: the goal is to help advance the hardware that makes powerful personal intelligence practical.
- Principle
- Compute beside memory. Activations move; weights stay.Design
- Runs
- A frontier-class model, the agent and Mynd, as one system, at homeDestination
- Why it pays
- A frontier mixture-of-experts model reads only a few percent of its weights for each token it writes, so speed comes down to memory bandwidth, not arithmeticThe model’s arithmetic
- Accuracy
- No accuracy lost: our expert engine computes a frontier model’s own 4-bit math exactly, matching it bit for bit on real weightsMeasured · in simulation
- Path
- FPGA, then a prototype pod, then custom hardware, then chip designThe route
To put a frontier-class intelligence in a home, we must solve three constraints:
- Bytes per token
Writing each token means reading gigabytes of weights, so speed is set by memory bandwidth: read less, or read it where the compute already is.
- Memory at home prices
A frontier model needs hundreds of gigabytes of fast memory at its released precision, and in 2026 it is memory, not compute, that sets the price.
- Agent-speed input
Agents read far more than they write: every tool result is thousands of new tokens, so compute has to sit beside every byte for reading as well as for writing.
The central idea
Mynd at every scale.
Mynd delivers truth by how memories and facts connect. The same mechanism scales: give each person in an organization their own agent and Mynd, give the organization a central Mynd, and connect them into one company mind that can be asked for exact, current truth about any task or plan, with every part traced to the people it came from.
The path
We will make that ambition concrete through successive demonstrations.
- 01 · Now
Axel, in invite-only beta
Mynd and the harness are live, and we’re taking on a small number of design partners for custom intelligence for organizations.
- 02 · Next
One million users
Our first major scale milestone. Our own inference on rented GPUs, and models adapted from strong open checkpoints.
- 03 · Then
The Axel pod
A prototype on available hardware, then custom hardware built around the model and its memory.
- 04 · Ultimately
Personal superintelligence
Chip design, personal AGI, and personal access to superintelligence, with privacy and ownership carried through.
The next frontier lab, made in Canada.
We’re a new company, and we know how far there is to go. We intend to think as big as possible and push our own limits, the company’s limits and the field’s limits.