Keyloh
In development

Telos systems.

Frontier-capable computing in forms that live where people work — quiet, compact, powered from the wall. Telos systems are being developed on commercially available accelerator and memory technologies, engineered around one decision: maximise unified memory, and let everything else serve it.

Reserve now The principles →

We are developing AI systems that address both memory-bandwidth-bound inference and compute-intensive model development — through unified memory, high-bandwidth data movement, workload-aware scheduling, and scalable multi-system configurations.

Engineered around memory.

A model must reside entirely in fast, inference-accessible memory to generate tokens at interactive speed. Every Telos decision follows from that constraint.

  1. Unified memory

    One pool, shared by CPU and accelerator — no partitions to manage, no copies to shuttle. The memory ceiling, not the marketing sheet, decides what a system can hold.

  2. Memory-local computation

    Work happens as close to the data as the platform allows — because moving tensors costs more than computing with them.

  3. Efficient model execution

    Serving tuned to the machine: models resident in memory, precision chosen per workload, multiple models served to whole teams at once.

  4. Modular scaling

    A system stands alone, or joins its full cluster and behaves as one pool — provisioned together, managed from one console.

  5. Software and hardware, co-designed

    Keyloh AI OS and the Telos runtime are built with the systems, not shipped after them — capability arriving ready rather than assembled. About the software →

Two configurations

Each machine comes standalone or as its full cluster — nothing in between. Interconnected and pre-configured before dispatch, a cluster behaves as a single pool of local intelligence.

Telos T128K
The deskside system — standalone, or a cluster of seven.Target specification: 256 GB of unified memory per system.
Telos M128K
The room-scale system — standalone, or a four-bay cluster chassis.Target specification: 891 GB of unified memory per system.
Target architecture
Target specificationUp to 3.5 TB of aggregate unified memory in a four-system configuration. Figures are engineering targets, not measured results; validated specifications are shared in the technical brief.
Serving
The largest openly available models, served locally — performance depends on model, precision, and context, and is quoted per configuration in conversation.
Operating environment
Pre-configured on Keyloh AI OS with the Telos Autotelic Runtime — models installed and serving from first boot.

Full engineering specifications are available in the technical brief, after qualification and where justified under NDA. Request it →

The largest open-weight models, entirely within your walls.

Concept imagery — research visualisation, not final hardware

Reservations are open.

A reservation is an enquiry, not a checkout — no payment is taken, and the specification is confirmed with you in conversation before anything is built.

Reserve now Ask a question →