Fresh heartbeat
A worker reports on a fixed interval. A snapshot older than the freshness window is not eligible for new work.
How it works
How a request moves
Two paths share one scheduler: one sends a request, the other supplies the machine that runs it. Access is invite-only for both.
Inference is the work of turning a prompt into output one token at a time. A worker is a contributor’s PC running the worker app and a local model backend. The scheduler matches each request to a worker that has the model loaded, enough capacity, and a fresh heartbeat.
Two paths
Two paths through the same scheduler.
Contribute capacity, or call the API. The same placement rules apply either way.
Both paths meet at the scheduler.
-
Check
Run the local check in your browser. It reads your graphics adapter or CPU, memory estimate, OS, and browser.
-
Get an invite
Access is invite-only. If the hardware fits, you receive an invite and a one-time enrollment token by email.
-
Install
Install the Windows worker app, enroll with the token, and confirm the model and backend it will serve.
-
Set limits
Choose the enable switch, VRAM budget, model-cache quota, idle-only wait, and active hours. The worker enforces them locally; CPU, power, thermal, and network ceilings are planned.
-
Go online
The worker registers its model and sends heartbeats. It accepts jobs only while every limit allows it.
-
Request access
Access is invite-only while the network grows. Request a key; it arrives when a slot opens.
-
Send a request
POST /v1/chat/completions with an OpenAI-compatible body. The live model key is lfm3.
-
Route
The scheduler filters candidates by model, heartbeat freshness, VRAM and RAM margin, and network limits.
-
Stream and account
Tokens stream back as server-sent events until [DONE]. Usage is accounted per request.
- Client
- Gateway
- Scheduler
- Worker
curl https://<your-domain>/v1/chat/completions \
-H "Authorization: Bearer $EASYAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "lfm3",
"messages": [
{ "role": "user", "content": "Summarize the lease rules." }
],
"stream": true
}'import os
from openai import OpenAI
client = OpenAI(
base_url="https://<your-domain>/v1",
api_key=os.environ["EASYAI_API_KEY"],
)
stream = client.chat.completions.create(
model="lfm3",
messages=[{"role": "user", "content": "Summarize the lease rules."}],
stream=True,
)
for chunk in stream:
print(chunk.choices[0].delta.content or "", end="")Placement
What the scheduler sees
Every placement is a filter, not a guess. A worker stays eligible only while all four checks pass.
Declared model and backend
The worker declares the model revision and backend it has ready. A request for a different model is not placed there.
VRAM and RAM safety margin
Placement keeps headroom so the model and context do not push the machine into swap or an out-of-memory stop.
Network limits
Worker network limits and schedule windows are checked before a job is assigned.
Failure handling
When a worker drops
A dropped worker is treated as a failed placement, not a failed request, until the bounded retry gives up.
-
Heartbeat
The heartbeat goes stale
The worker misses its report window and becomes ineligible for new placements.
-
Lease
The lease still holds
The in-flight request keeps its lease while the scheduler checks whether the worker is still responding.
-
Replan
Bounded retry
If the lease fails, the scheduler retries on other eligible workers, with at most two attempts.
-
Outcome
Clean failure
If no attempt succeeds, the request fails with an error. A partial stream is never presented as complete.
Current limits
What the network can’t do yet
The network is small and deliberately limited. These are current constraints, not future promises.
Windows only
The worker app runs on Windows 10/11 64-bit. macOS and Linux workers are not supported.
LFM3 only
LFM3-1.7B is the only model supported by the current worker build. Qwen3 models are planned, not callable.
Invite-only access
There is no public signup. API keys and contributor invites are reviewed manually.
No phones
Phones and tablets cannot contribute capacity, and there is no mobile app.
Pipeline parallelism in progress
A request runs on one worker. Pipeline parallelism across machines is in progress and not available.
Credits not live
Contributor credits are not live yet. Terms are set per invite while the network grows.
Next step
Start with the path that fits.
Check what your PC can host, or request API access and we will email a key when a slot opens.