The world's largest peer-to-peer AI network.
Every device is a node. Every node makes the network smarter.
There are billions of devices sitting idle right now — laptops, desktops, gaming PCs, workstations, even Raspberry Pis — with CPUs and GPUs doing nothing. Meanwhile, running AI costs a fortune in cloud compute, and access is controlled by a handful of companies.
HiveBear connects these idle devices into a single distributed AI network. When you join the mesh, your hardware contributes to a collective compute pool. When you need to run a model that's too large for your machine, the mesh splits it across multiple devices automatically. No central server. No cloud bill. No data leaving the network.
The goal is simple: build a global P2P mesh where anyone can run any AI model, regardless of what hardware they own, by pooling compute with everyone else.
You (8GB laptop) Friend (16GB desktop) Mesh peer (GPU workstation)
| | |
+------------- QUIC/TLS encrypted mesh ---------------+
|
HiveBear Mesh Network
|
Distributed inference: 70B model
split across all three devices
- Install HiveBear on any device
- Join the mesh — your device auto-profiles its hardware and advertises its capabilities
- Run any model — if it fits locally, it runs locally. If it doesn't, HiveBear distributes the model layers across mesh peers automatically
- Contribute idle compute — when you're not using your device, it helps others run their models
# Join the global mesh
hivebear mesh start
# Run a 70B model you couldn't run alone
hivebear mesh run llama-3.1-70b --prompt "Explain quantum computing"
# See who's connected
hivebear mesh statusThe mesh uses QUIC transport with TLS encryption. Inference is distributed using pipeline parallelism — each device holds a subset of model layers and forwards activations to the next peer. No raw model weights or user prompts are exposed to other nodes.
Even without the mesh, HiveBear is a complete local AI runtime. It profiles your hardware, picks the best model and quantization automatically, and runs it:
# One command: profile hardware, pick best model, download, chat
hivebear quickstart
# Or use it as an Ollama-compatible API server
hivebear serveThe serve command is a drop-in replacement for ollama serve — same port (11434), same API. Your existing tools, IDE extensions (Continue, Cody), and scripts work without changes. When a model is too large for your hardware, it automatically overflows to the mesh.
# One-line install (Linux/macOS)
curl -fsSL https://raw.githubusercontent.com/BeckhamLabsLLC/HiveBear/main/install.sh | bash
# Homebrew
brew install BeckhamLabsLLC/hivebear/hivebear
# Scoop (Windows)
scoop bucket add hivebear https://github.com/BeckhamLabsLLC/scoop-hivebear
scoop install hivebear
# Build from source
cargo install --git https://github.com/BeckhamLabsLLC/HiveBear hivebear-cli --locked
# --locked builds the dependency versions this release was tested with.
# Without it cargo re-resolves every dependency to the newest compatible
# release, which is not what we build or test.Docker images are not published yet. The
build-dockerjobs failed on every release up to 0.1.6, because the build images lacked the libclang that llama-cpp-sys-2's bindgen step needs. That is fixed, but a package pushed to ghcr for the first time is private, soghcr.io/beckhamlabsllc/hivebearis not pullable until it is made public. These instructions will return then.
HiveBear auto-detects and adapts to whatever you have:
| Device | RAM | Solo | With Mesh |
|---|---|---|---|
| Raspberry Pi 5 | 8 GB | TinyLlama 1.1B, Phi-2 2.7B | Contribute layers to larger models |
| Old laptop | 8 GB | Llama 3.1 8B (Q4), Mistral 7B | Help run 13B-30B models |
| Gaming PC | 16 GB | Llama 3.1 8B (Q8), CodeLlama 13B | Help run 70B+ models |
| Workstation | 32+ GB | Llama 3.1 70B (Q4), Mixtral 8x7B | Run anything |
GPU acceleration is automatic (CUDA, Metal, Vulkan, WebGPU).
Rust workspace, 8 crates:
hivebear-core Hardware profiling, model recommendations
hivebear-inference Multi-engine inference (llama.cpp, Candle)
hivebear-mesh P2P distributed inference over QUIC/TLS
hivebear-registry Model search, download, conversion (HuggingFace)
hivebear-persistence Conversation history (SQLite)
hivebear-cli CLI + API server (Ollama + OpenAI compatible)
hivebear-web WASM bridge for browser inference
apps/desktop Tauri desktop app (Rust + React)
hivebear quickstart Profile -> recommend -> install -> chat
hivebear serve Start Ollama + OpenAI compatible API server
hivebear profile Show hardware capabilities
hivebear recommend Get model recommendations for your hardware
hivebear mesh start [--port 7878] Join the P2P mesh network
hivebear mesh status Show connected peers and network capacity
hivebear mesh run <model> Distributed inference across the mesh
hivebear mesh stop Leave the mesh
hivebear search <query> Search models on HuggingFace
hivebear install <model> Download a model
hivebear run <model> Local inference (chat, --api, or --prompt)
hivebear list / remove / storage Manage installed models
- CLI: Linux, macOS, Windows, ARM (Raspberry Pi, Apple Silicon)
- Desktop app: Linux (.deb, .AppImage), macOS (.dmg), Windows (.msi, .exe)
- Mobile: Android (.apk)
- Browser: WASM + WebGPU
- Docker: CPU and CUDA images
Official HiveBear builds send anonymous crash reports, so we hear about the things that break on hardware we do not have. It is on by default and takes one command to turn off:
# Any of these disables it permanently
export HIVEBEAR_TELEMETRY=0
export DO_NOT_TRACK=1 # honoured, as a matter of course
hivebear sentry-check # show whether it is on, and whyIn the desktop app it is a toggle under Settings → Crash Reports.
What is sent: the error and its stack trace, the HiveBear version, your OS and CPU architecture, and a random identifier generated on your machine.
What is never sent: your prompts, chat history, model files, file contents, email address, IP address, API keys, or mesh identity. File paths are stripped of your username before they leave the machine.
The random identifier exists only so we can tell "one person hit this 400 times" apart from "400 people hit it once". It is not derived from your hardware or your mesh keypair, and it is not linked to anything else.
If you build HiveBear from source, there is no reporting endpoint compiled in at all — a source build cannot report anything even if the setting says it should. That includes every distro package built from this repository.
See CONTRIBUTING.md. The most impactful contributions right now are around the mesh networking layer and hardware profiling coverage.
MIT. See LICENSE.