How SwarmLLM Compares
SwarmLLM sits beside two kinds of project: apps that run AI on ONE machine (Ollama, LM Studio, Jan), and networks that split a model across MANY (Petals, exo, Bittensor). This page compares it with the second kind, where the differences are less obvious.
| SwarmLLM | Petals | exo | Bittensor | |
|---|---|---|---|---|
| Install | Download and run (one program) | pip install | pip / source / macOS app | pip + blockchain setup |
| Language | Rust | Python | Python | Python + Substrate |
| Where the computers are | Internet, your home network, or your own devices — found automatically | Internet (volunteers) | Your home network or Tailscale | Internet (blockchain) |
| Traffic between computers | Encrypted (libp2p Noise / TLS; activations additionally sealed with X25519 + ChaCha20-Poly1305) | Encrypted (libp2p) | — | Subnet-dependent |
| Can the computers doing the work read your request? | Yes, in the public swarm — so SwarmLLM has Private Mode, which keeps requests on your own devices (and, by default, other SwarmLLM computers on your local network) | Yes (Petals' own wiki) | Yes (they are yours) | Yes |
| Rewards | None — no token, no payment, no blockchain | Name on a monitor page | None | TAO token (real money) |
| How a model is split | Pipeline (tensor parallelism exists for fast local networks but is off by default) | Pipeline | Tensor + pipeline | Subnet routing |
| Model families verified on real models | Llama, Qwen 2/3, Qwen 3.5 (dense), Mistral, Gemma 2, Phi-3/4, GLM-4 (checked at every release); Gemma 3 (text) checked against llama.cpp on the 1B and 4B, Gemma 4 (text) on the 12B, the E2B, the E4B and the 26B-A4B mixture of experts. Llama 4 and Qwen 3.5's mixture-of-experts models are checked against llama.cpp on small models only; StarCoder2 and DeepSeek are not supported yet | Llama, Mixtral, Falcon, BLOOM | Llama, Mistral, Qwen, DeepSeek, LLaVA | Any (subnet-defined) |
| Needs the whole model file? | No — each computer downloads only its parts | Loads whole blocks | Yes | N/A |
| Cloud fallback | Optional, 12 providers with your own keys | No | No | No |
| Images and adapters | Vision models (LLaVA verified) and per-request LoRA on a model held whole | LoRA | Vision experimental | Subnet-specific |
| API | OpenAI + Anthropic + MCP (works as a Claude Code backend) | PyTorch / Transformers | OpenAI + Claude + Ollama | Subnet-defined |
| Built-in app | Dashboard, chat, setup wizard; 21 languages | Basic chatbot | Dashboard | None |
| SDKs | Python and JavaScript clients in the repo (with LangChain and LlamaIndex adapters); not yet on PyPI/npm | Python | — | Python |
⚠ Splitting a model is only fast when the computers are close. Every network on this page passes your request through each computer for every word of the reply, so the round trip between them is paid once per word. Two computers on the same local network are fast; two on different continents are slow. SwarmLLM's plan for automatically forming teams of nearby computers is in the regional pipelines plan.
Details on the others change quickly — check each project's own page before relying on a row.