Unlimited access to advanced AI
Inference runs on the hardware you already ship, resulting in no per-use charges and no costs that scale with usage.
The world’s first on-device AI that runs on any hardware
Runs on low-power hardware ｜ Cloud-level intelligence ｜ Works offline ｜No privacy risk ｜ Zero token cost
Explore our models
Designed for low-power, low-memory devices
Up to 3 GB of memory, bringing powerful AI to compact hardware from phones to consumer-grade boards.
Private by design
Keep sensitive data secure with local processing that keeps data on your device.
How far down the device ladder
can on-device AI actually reach?
Most stay at the research stage
A few reach high-compute hardware
A handful stop at mid-compute hardware
Edge0All Four Tiers
Default state: no stage gradient background shown.
Parameter count is not the same thing as RAM
See the modelsEdge0-35B · sparse MoE
1.06 GB
LFM2-2.6B
5.4 GB
MiniCPM-V 4.0
9.0 GB
Gemma 4 E2B
11.4 GB
MiniCPM4-8B
16.4 GB
Gemma 4 E4B
17.9 GB
Qwen3.8-27B
56.0 GB
Muse Glimmer 30B
56.0 GB
Qen3.5-35B-A3B
70.0 GB
DeepSeek-R1-Distill-Llama-70B
140.0 GB
A family of models.
One engineering discipline.
All models and specs
Edge0-35B
A 35-billion-parameter flagship on-device model that runs with 1–3 GB of memory and supports understanding and execution of complex tasks.
Edge0-8B
An 8-billion-parameter lightweight on-device model that runs with 1–2 GB of memory, suitable for Q&A and low-to-medium complexity tasks.
Audio8-ASR-3B
On the open-source ASR leaderboard — world-best 5.04% WER
Audio8-TTS-0.6B
0.6B size, 8B-grade voice — zero-shot cloning that captures the exact texture of your voice
Silicon, cloud and the companies putting models into products
//2026
Smart Earbuds
Edge-cloud voice SDK embeds in earbuds: natural commands, live readouts, zero hardware changes.
//2026
Robotics Chip
Edge-cloud HRI SDK embedded in humanoid robots — natural-language dialogue and command execution, fully functional offline or on weak networks.
//2026
Smart Glasses
Edge-cloud coordination cuts latency and cost: frequent commands on-device, complex tasks in cloud.
//2026
Embodied AI
Robot chips plus edge voice models: out-of-the-box, offline, low-cost voice interaction for makers.
//2025
Smart Toys
Edge-side models on a hybrid edge-cloud base answer fast commands and recognize emotion.
AI that consumer hardware can actually run.
Inference happens on hardware you already shipped, so there is no token meter to run down and no bill that grows with usage.