glassfire pre-release

AI that lives on the device  { −1 · 0 · +1 }

Reasoning per joule.

The reasoning brain for machines that can't depend on the cloud — robots, drones, vehicles, smart devices. Glassfire runs looping, ternary-weight models entirely on-device, with deeper thinking on demand: trade latency for reasoning depth, per request. First light soon.

  • On-device
  • Works offline
  • Private
  • Low-cost

The runtime is built and working today — on Apple silicon and NVIDIA GPUs, faster than the best open-source engine with identical outputs, serving looped reasoning models that stock engines can't run.

  • 1.58 bits / weight
  • bit-exact
  • depth on demand
  • Metal
  • CUDA
  • NEON
  • CPU
  • WASM
  • NPU

Benchmarks, releases, early access. A few emails a year — no noise.

The glass: ternary weights are immutable and deterministic — you can attest and audit exactly which model produced every answer.