Questions, answered plainly

No. LFM3-1.7B can run on a CPU with 8 GB of RAM and 4 or more cores. A GPU with at least 4 GB of VRAM usually produces output sooner. The browser check estimates a tier; the Windows app measures the machine before it can accept jobs.

See the requirements

A Windows laptop that meets the requirements can contribute. Heat, battery, and power limits matter, so use the app’s pause, idle-only, and active-hours controls. The browser estimate is rough; the app confirms the machine.

The worker stops sending heartbeats and receives no new jobs. A request already running may fail or be retried on another eligible worker, with at most two attempts. If no attempt succeeds, the request returns an error; a partial stream is never presented as complete.

Yes. The app lets you enable or disable contribution, set idle-only minutes and active hours, declare a VRAM budget, and set a model-cache quota. The current desktop build enforces the enable switch, idle-only setting, and active hours locally. CPU, GPU, and bandwidth ceilings are in the design but are not exposed in this release.

Read about worker limits

Each PC enrolls separately with its own invite and one-time enrollment token. Invites are reviewed manually, so request one for each machine you want to add.

Uninstall the app from Windows Settings → Apps. Uninstalling removes local data from that PC but does not remove the worker from the network. Choose Revoke in the app while it is running to remove the worker; if the app cannot reach the server, pause it and retry revoke later.

Pause, resume, and revoke

Windows 10 or Windows 11, 64-bit. The current worker build does not support macOS, Linux, phones, or tablets.

Only LFM3-1.7B is live in this release. The scheduler adds a 15% margin over the model plan’s VRAM requirement, so a stage that needs 9,500 MB requires at least 10,925 MB of declared or detected VRAM. The invite names the model and role before you install.

The worker loads the model into the VRAM budget you declare and runs only the jobs it accepts. Games and other GPU work share the same card and may slow down while inference runs. Pause the worker when you need the machine.

The current requirements are written as 4 GB of VRAM or a CPU path, without a vendor list. AMD, Intel, and integrated GPUs depend on the driver and the backend in use, so the app checks the actual machine after install rather than promising a model will load.

Access is invite-only in this release. There is no self-serve plan, no public signup, and no card form. Credits and pricing are not live. If an allocation is available for your invite, the operator will state it in the invite.

Request access from the contribute page. The operator reviews requests manually and emails a key when a slot opens. There is no public signup, login, or key-creation endpoint.

Request access

The chat endpoint accepts an OpenAI-style chat body at /v1/chat/completions, including messages, max_tokens, and stream. It returns a chat.completion object or server-sent chunks. Compatibility is limited to the documented endpoints and fields; this is not an OpenAI partnership.

Chat completions reference

LFM3-1.7B is the only live model, with the model key lfm3. Qwen3 models are planned and cannot be called. The gateway limits a key to 60 requests per minute and rejects an estimated input above 2,048 tokens. Output is capped at 4,096 tokens per request; the default is 512.

See the models

The scheduler holds a short lease. If the worker drops, it retries on other eligible workers, with at most two attempts. If no attempt succeeds, the request fails with an error such as NO_CAPACITY or DEADLINE_EXCEEDED. A stream without the [DONE] sentinel is incomplete.

Read the error codes

Credits are not live in this release. We won’t publish rates until they are.

No user-facing credit balance or payout is live in this release. The gateway still accounts for usage internally, but there are no published rates, no contributor payout balance, and no public balance page.

We can’t give you a number. It depends on your hardware, the work your PC does, and your local electricity tariff. Inference draws more power than an idle PC, and you cover that cost. The app does not measure or bill electricity.

There is no payout program in this release. No contributor is owed credits, revenue, or any other payment. If the operator introduces one, the terms will be published before anyone opts in.

Prompt text is not written to Postgres or server logs by default. For a worker-executed job, the gateway holds the prompt briefly in Redis, encrypted at rest in the hosted service, and deletes the hold when the job settles. Completion text is streamed to the client and its per-job Redis stream is deleted when the job settles. Sampled verification copies are deleted when the verdict is recorded.

Privacy notice

The assigned worker sees the prompt and intermediate inference data. No prompt logging by default; the service is not end-to-end private. That worker runs the model on its own PC, so treat prompts as visible to the operator of that machine.

No. Inference is not end-to-end private. The worker that runs a request can inspect the prompt and the intermediate data it processes. Encryption in transit protects the connection, not the machine at the other end.

Workers enroll with a one-time token and use a short-lived session token on every control-plane call. The operator excludes the source worker when a completed job is sampled for re-execution on a second worker and compares the two outputs. Repeated mismatches quarantine the worker from new placements. Verification runs after delivery, so it catches some bad output rather than preventing every bad answer.

Security model

No. Phones and tablets cannot run the worker or contribute capacity. The browser check explains this when it detects a mobile device.

The worker app runs on Windows 10/11 64-bit only. macOS and Linux are not supported in this release.

No mobile app is available. The qualification check runs in a browser, but capacity contribution requires a Windows PC.