Helix-M is a palm-sized personal AI appliance powered by a single Tsavorite OPU. It delivers 1,000 TOPS of AI performance with up to 96GB of unified high-bandwidth memory in a mini form factor. It runs multimodal models, inference, and fine-tuning locally, so you get private, low-latency intelligence on your desk without depending on the cloud.

Helix-M runs models up to ~100B parameters locally, including large language, vision, and multimodal models. It supports popular runtimes and tooling such as Ollama and llama.cpp, so you can deploy open models and end-to-end AI workflows with minimal setup.

Because every token is processed on-device, your data never leaves the system. There are no cloud round-trips and no third-party inference services in the loop. That makes Helix-M well suited to healthcare, legal, industrial, and other environments where data sovereignty and confidentiality are non-negotiable, and it can operate fully offline.

Helix-M needs no special power, cooling, or networking. It is an all-inclusive appliance that does not need any additional monitors, keyboard etc and works with existing phone/laptops/desktops etc.

Helix-M delivers PetaFLOP-class performance in a fraction of the power and footprint of a traditional GPU workstation, with a large unified memory space that avoids splitting models across devices. For teams that need more capacity, Helix-D and rack-scale Helix-R scale the same architecture up to larger models and higher throughput.