What we open source, and what we don’t
Sending your data to someone else’s GPUs means trusting them. The interfaces, protocols, and reference implementations are open so you can check how it works and build against it, and the list below says plainly what stays closed and why. When it is a judgement call, we open it.
View the code on GitHub →What we open source
- The OpenAI-compatible API surface and a reference server — backed by a conformance suite that drives it with the official OpenAI SDK. view repo → Available
- The capacity-provider interface and a complete AWS Spot reference implementation, with a conformance suite. view repo → Available
- The payload-encryption format — the code that encrypts your data at rest — so the privacy claim is auditable. view repo → Available
- A dependency-free Hugging Face Hub client for model discovery. view repo → Available
- The worker agent — the exact code that runs on a compute node and touches your data. Planned
- The model-runtime interface and a vLLM reference, plus reproducible worker image builds. Planned
- Reproducible benchmarks and the methodology behind every cost/quality/latency number we publish. Planned
- A public, “Steam hardware survey”-style dataset on models, instance types, and real-world cost. Planned
Open components are Apache-2.0. Model weights are governed by their own upstream licenses (Qwen, Gemma, GLM…), which we respect and surface per model.
What stays proprietary (for now)
- The tuned speculative pricing and placement heuristics — our actual arbitrage edge.
- Model fine-tunes and optimization recipes that lower our cost basis.
- Commercial machinery: billing, accounts, settlement, fraud, reputation.
Nothing security-relevant to your data hides behind the proprietary line. The data path (payload encryption) and the API protocols are open and conformance-tested today, and the worker agent is next, precisely so the privacy claims are ones you can check.
What the open code lets you check
Auditable data path
The payload-encryption format — the code that encrypts your data at rest — is open (the worker agent is next), so “no payload logging, scoped credentials, per-job encryption” is checkable, not asserted.
Reproducible numbers
Every cost, quality, and latency claim ships with the code and method to reproduce it. We’d rather publish a modest number you can verify than an impressive one you can’t.
Conformance-tested compatibility
The OpenAI-compatible surface is tested against the official OpenAI client libraries, in CI, on every merge. Compatibility is backed by a suite that runs rather than by a bullet point.
Read the code, then send us a batch.
The API surface, the capacity interface, and the data path are open — with conformance suites. Leave your email and we’ll invite you to the alpha.
No credit card, no spam — one email when your invite is ready.
Closed alpha — onboarding is gated while we calibrate. Already invited? Sign in.