Edge0

On-device AI built for consumer hardware
Cloud-level AI · Offline access · Strengthened privacy · No pay-per-use fees

Website · Models · GitHub · X

Edge0 builds efficient foundation models for phones, laptops, wearables, robots, and other consumer hardware. Our models are designed for local, private, and offline inference.

LLMs

Latest release

• Edge0-145B · Q3_K · NVFP4
  145B latent MoE for local coding and agent tasks on consumer hardware, trained natively at 4 bits.
• Edge0-120B · Q3_K · NVFP4
  120B latent MoE for local coding and agent tasks, with smaller weight files and native 4-bit training.

Previous release

• Edge0-35B-A3B — 35B sparse MoE with about 3 GB peak active memory.
• Edge0-8B-A1B — lightweight 8B sparse MoE with about 1 GB peak active memory.

Speech models

• Audio8-ASR — 4B Infinite · 3B · 0.6B · 0.1B — speech recognition from flagship to edge tiers.
• Audio8-TTS — Realistic Air · 0.1B — multilingual, zero-shot voice generation with voice cloning.

Explore all models on Hugging Face, and the open-source inference stack on GitHub.

Research · About · Contact