AI Internetalpha-q3-2026
Distributed network for running large language models on a community of home GPUs. A Cursor alternative without subscription — your code stays on your perimeter.
Screenshots
Documentation
What it is
AI Internet is a distributed network for running large language models on a community of home GPUs. A Cursor and Copilot alternative that runs on your video card with no subscription. Simple queries are answered by a local model instantly. Complex ones are assembled by an ensemble of five to seven specialised models.
How it works
Each node hosts its own specialist: DeepSeek-Coder for programming, Qwen-Math for math, a Russian-language specialist. The Mixture of Agents approach (Wang et al., 2024, arxiv:2406.04692) — an ensemble of six small models beats GPT-4o.
Security
Channel between nodes is encrypted: Noise XX over QUIC and TLS 1.3, X25519 + Ed25519 keys, ChaCha20-Poly1305. Session keys rotate every 5 min. 5–7 independent BFT coordinators, signed DHT, geographic peer diversity. Executable code cannot transit the network — only numbers and tokens flow.
Roadmap
- Months 1–3. Prototype distributed inference, basic verification, ten-node testnet.
- Months 4–6. Closed alpha: Code Agent, VS Code extension, sandbox, 50 nodes, 200 users.
- Months 7–12. Open beta: public reputation log, partner node pool.
- Months 13–24. Maturity: shared memory, tool marketplace, Tor mode.