Interactive web experiences, generated on-device. Edge0 can turn natural-language requirements into frontend structure, logic, and interactions, then iterate on the result locally. From the first prototype to the final refinement, the workflow stays on-device.

3D scenes, built locally. Describe what you want in plain language. Edge0 can generate a structured 3D scene, then modify and refine it on your phone or PC. No project data leaves the device.

What can you build with Edge0? Playable games, entirely on-device. Start with a natural-language idea. Edge0 can generate the game logic, write the code, and refine the result locally. Around 2 GB of peak memory. No cloud calls.

How does Edge0 run a 35B model with only 1–2.5 GB of peak memory? In simple terms: • The full model stays in storage. • Only the parameters needed for the current step are loaded into memory. • The next parameters are prepared while the current step is running. • Memory is reused instead of loading the entire model at once. Large model. Small working set. Fully on-device.

Edge0 is now open source. The first release includes 35B and 8B models. It also includes the runtime and inference code to run them locally. Runtime + inference code: Models: ⚡️ 8B: 🧠 35B: Built for phones, laptops, PCs, and robots. Fast, private, and fully on-device.

A 35B language model running on an iPhone using only 1–2.5 GB of peak memory. No cloud. No remote server. No desktop GPU. Today, we’re open-sourcing Edge0 — a framework for running large AI models fully on-device.

Edge0 Technical Breakdown Edge0 is an open-source, streaming MoE inference framework for running large models on devices with tight memory budgets. 💾 SSD expert offload: MoE expert weights are memory-mapped on SSD and streamed on demand. Hot experts stay cached in memory, while long-tail experts are prefetched layer by layer. 🔮 Prerouter routing prediction: Edge0 predicts the next token’s expert routes one step ahead, overlapping SSD reads with the current forward pass. 🧩 Parallel LoRA: LoRA adapters are applied at runtime rather than merged, so a single read-only base model can serve multiple adapter sets. Put simply: the full model lives on storage; only the weights needed for each step are in active memory. Report：

Updated technical report:

I need a prompt for this. 🔥