Hellas runs AI jobs on GPUs you don't own, and on your own GPU for people you trust. Your machine renders each job, for example a Whisper transcription or a Qwen chat turn, into an exact program with content-addressed inputs. The provider runs that program and signs the result. Your machine checks the signature and decodes the result. You pay a provider with devnet coins, run under access someone shared with you, or run the job yourself.
curl -fsSL https://github.com/hellas-ai/hellas/releases/latest/download/install.sh | shThis installs the portable client. GPU machines run the ghcr.io/hellas-ai/hellas:hip
or :cuda image, or the NixOS module.
Hellas does not host model weights. Every machine converts the same pinned checkpoint, and a pinned commit converts to the same bytes on every machine. That is how your job names the provider's weights by hash. Fetch Whisper tiny, a small Qwen3 and a sample recording:
hellas convert-whisper openai/whisper-tiny@169d4a4341b33bc18d8881c4b69c2e104e1cc0af -o whisper-tiny
curl -fLo qwen3.gguf https://huggingface.co/unsloth/Qwen3-0.6B-GGUF/resolve/50968a4468ef4233ed78cd7c3de230dd1d61a56b/Qwen3-0.6B-BF16.gguf
curl -fLo speech.wav https://raw.githubusercontent.com/ggml-org/whisper.cpp/b0a11594aec50892a02cd8d129eee2dfe93a8bb8/samples/jfk.wavOn a GPU machine, one command serves both packages. Qwen uses GGUF; Whisper and Flux use the directories produced by their converters. Hellas detects the kind and prints each model, API path and backend at startup.
hellas gateway --model ./qwen3.gguf --model ./whisper-tiny --wrap claude--wrap gives the child its OpenAI and Anthropic URLs and a private bearer,
then drains the gateway when the child exits. Without --wrap, the gateway
runs in the foreground and prints its URL and bearer to the terminal. For
headless use, --bearer-file PATH writes the bearer atomically with mode 0600;
without a wrapper, file or terminal, startup fails. Routing configuration is
not saved, and each start creates a new bearer. Repeatable service deployments
use the shared --config TOML
with gateway or serve; NixOS settings generate the same file.
An upstream works without a GPU. The key file is read at startup:
hellas gateway \
--upstream openrouter=https://openrouter.ai/api,key-file="$HOME/or.key" \
--model gpt-5@openrouter --wrap claudeRepeat --model NAME@a --model NAME@b to specify failover order. Several
models and backends can serve requests concurrently. A request can change
backends only before a proposal; an accepted job is never replayed on another
backend. See the HTTP gateway guide.
Start the persistent daemon on the GPU machine:
hellas serveA new node prints one hellas-invite: line and waits for its owner. In another
shell on that machine, set INVITE to that full line, claim the node and add
the packages already stored there:
hellas pair "$INVITE" --name mine
hellas route add "$PWD/whisper-tiny"
hellas route add "$PWD/qwen3.gguf"
hellas routesEach edit is one atomic configuration change. The daemon restores the saved configuration after restart. It publishes bounded resources for the packages; sharing uses the existing authority and allowance commands.
Your friend runs hellas contact and sends you their contact. Set
FRIEND_CONTACT to that full contact string. Give that
contact ten requests to the qwen3 resource for a day and send them the
resulting access bundle. Select the resource explicitly when the node serves
several packages; hellas routes lists their names:
hellas share --to "$FRIEND_CONTACT" --resource qwen3 --limit requests:total:10 --expires 1d --out friend.accessYour friend imports it under a contact name and renders requests from their own copy of the same package or a metadata-only descriptor:
hellas pair friend.access --name alice
hellas gateway --model ./qwen3.gguf@alice --wrap claudeEach native job runs in its own process group. This contains crashes and execution timeouts; it is not a security sandbox. Programs can access the provider user's files and devices. Only grant execution to people whose programs you trust. See node administration.
The same packages work with the job commands. On your GPU:
hellas transcribe --model whisper-tiny --audio speech.wav --localOr use a friend's access bundle and their GPU evaluator:
hellas transcribe --model whisper-tiny --audio speech.wav --access friend.access \
--backend hellas/catena-dfad87438018-hipA paid job uses devnet coins and checks the quoted price before signing:
hellas faucet
hellas transcribe --model whisper-tiny --audio speech.wav --pay 20 \
--provider "$DEVNET_PROVIDER" --backend hellas/catena-dfad87438018-hip--local is available in GPU builds. NVIDIA providers use the CUDA evaluator.
See jobs for text, image, Fetch and prepared-program commands.
- HTTP gateway: models, upstreams, sessions and cancellation.
- Node administration: sharing, renewal, revocation and authority.
- Selling on the devnet: paid terms, stake and channels.
- Chain: querying the devnet and running validators.
- Nix, Docker and NixOS: builds, modules, images and development checks.