Evidence needed
Minimum
12 GB to under 20 GB VRAM
0/1 accepted evidence sets
Worker software
AI Power Grid Workers
Keep the models and runtime you already operate. Add a Grid worker beside them, choose what capacity to expose, and serve compatible jobs from the network.
Linux text workers are available now. Image, video, and audio onboarding is in qualification.
Connect your backend
Choose your operating system
Final compatibility is checked locally.
Before you download
- Backend
- Ollama or a tested compatible API
- Platform
- Linux · verified artifact
- Compatibility
- The selected model must be served by your backend
- Controls
- Model, limits, concurrency, schedule, pause
- Network need
- gpt-oss-120b: 1 serving, target 3; 3,491 completed jobs in the last 30 days.
- Worker sees
- Plaintext request inputs and outputs
- Rewards
- Accepted work contributes to the current AIPG period split
- Maturity
- Public text worker
The verified installer detects Linux x64 or ARM64, checks the release manifest and binary, and installs without starting the worker. Your inference backend remains a separate local service.
Join the operator cohort for setup supportFirst run on Linux
- 1.Start the Ollama or OpenAI-compatible backend that already serves your model.
- 2.Run the checksum-enforcing installer. It installs the worker but does not start it:
cd ~/Downloads chmod +x install-worker.sh ./install-worker.sh ~/.local/bin/grid-inference-worker --verify-runtime ~/.local/bin/grid-inference-worker - 3.Complete the local wizard at
http://localhost:7861. Version 0.3.8 uses a scoped Grid API key from the developerConsole. Enter it only in the local wizard, never in a shell command or public issue. - 4.Wait for the dashboard to report Online, then confirm the exact worker name or ID with the public worker check below.
The worker never needs a wallet private key. Configure the payout wallet separately in Console settings after the worker is healthy.
Bring your own runtime
Keep the AI stack you already run.
The worker runs beside your inference service. It does not replace your runtime, upload your model files, or silently advertise every model it discovers.
Compatibility
What can connect today
OllamaOpen
Text · Detected automatically
vLLM / LMDeployOpen
Text · OpenAI-compatible endpoint
SGLangOpen
Text · OpenAI-compatible endpoint
LM StudioOpen
Text · Detected or entered locally
KoboldCppOpen
Text · Detected or entered locally
Other OpenAI-compatible APIsAdvanced
Text · Operator-entered endpoint
ComfyUIQualification
Image / video · Reviewed model and recipe profiles
ACE-StepQualification
Audio · Reviewed direct-runtime profile
Open means operators can connect through the current public text worker. Qualification means the runtime needs a reviewed Grid profile and canary before it may advertise that capability.
You choose capacity
Text operators control the advertised model, response limits, concurrency, and operating schedule. Pause without uninstalling.
Your model stays yours
Model weights stay in your runtime. The worker needs a scoped Grid credential, never a wallet private key.
Workloads are not private yet
Community workers process plaintext prompts and outputs. Do not send secrets or regulated data until confidential execution is independently verified and available.
Operator cohort
Add independent text capacity
The current paid cohort is open to Linux operators with an existing Ollama or OpenAI-compatible GPU backend. Join the tracked cohort for current model gaps, acceptance evidence, and setup support.
Both GitHub paths are public. Share coarse hardware and availability only. Never post credentials, wallet details, network addresses, or private logs.
Operator planner
Find the useful path for your machine
Choose the workload you want to offer and adjust the example specs to match your machine. Nothing is submitted. Your local backend must still pass a generation test.
Likely starting point
Start with the text worker and your existing backend
Start Ollama or an OpenAI-compatible local server, load a model that fits your hardware, then select and test it in the worker. VRAM alone cannot determine model fit: quantization and context length also matter.
A verified worker download is available for your selected platform. Check current network needs separately; they are not recommendations for what fits your GPU.
View worker downloadsNetwork-priority text route
gpt-oss-120b
1 serving worker and 3,491 jobs producing 1,445,577.28 accepted den in 30 days. Priority uses accepted den and missing replicas, not raw request count or hardware compatibility. Advertise it only when your backend genuinely serves that model.
- Entered accelerator
- NVIDIA
- Expected text speed
- Not entered
See current capacity needs and workload history
Live network opportunity
Capacity and actual 30-day work
Jobs per worker is a rough workload-share signal, not a payout forecast. Throughput is observed across the existing Grid and is not a benchmark for your GPU.
Smollm-135m
text
- Workers
- 1
- Jobs, 30d
- 57,831
- Jobs / worker
- 57831.0
- Observed
- 396.3 tok/s
deepseek-v4-flash-nvfp4
text
- Workers
- 1
- Jobs, 30d
- 4,267
- Jobs / worker
- 4267.0
- Observed
- 88.8 tok/s
gpt-oss-120b
text
- Workers
- 1
- Jobs, 30d
- 3,491
- Jobs / worker
- 3491.0
- Observed
- 50.3 tok/s
qwen38-flash-next-125b-nvfp4
text
- Workers
- 1
- Jobs, 30d
- 493
- Jobs / worker
- 493.0
- Observed
- 6.3 tok/s
qwen3-27b
text
- Workers
- 1
- Jobs, 30d
- 440
- Jobs / worker
- 440.0
- Observed
- 82.7 tok/s
FLUX.2 Klein 4B FP8
image
- Workers
- 1
- Jobs, 30d
- 79
- Jobs / worker
- 79.0
- Observed
- 4.1s average
LTX-2.3
video
- Workers
- 1
- Jobs, 30d
- 56
- Jobs / worker
- 56.0
- Observed
- 67.9s average
gpt-oss-20b
text
- Workers
- 1
- Jobs, 30d
- 37
- Jobs / worker
- 37.0
- Observed
- 157.4 tok/s
Krea 2 Turbo
image
- Workers
- 1
- Jobs, 30d
- 31
- Jobs / worker
- 31.0
- Observed
- 8.6s average
ace-step-v1.5-xl-turbo
audio
- Workers
- 1
- Jobs, 30d
- 25
- Jobs / worker
- 25.0
- Observed
- 14.1s average
z-image-turbo
image
- Workers
- 1
- Jobs, 30d
- 23
- Jobs / worker
- 23.0
- Observed
- 18.8s average
LTX Director 2.0
video
- Workers
- 1
- Jobs, 30d
- 0
- Jobs / worker
- 0.0
- Observed
- No recent timing sample
LTX-2.3 Audio
video
- Workers
- 1
- Jobs, 30d
- 0
- Jobs / worker
- 0.0
- Observed
- No recent timing sample
| Model | Type | Workers | Jobs, 30d | Jobs / worker | Observed | Resilience |
|---|---|---|---|---|---|---|
| Smollm-135m | text | 1 | 57,831 | 57831.0 | 396.3 tok/s | Single-worker risk |
| deepseek-v4-flash-nvfp4 | text | 1 | 4,267 | 4267.0 | 88.8 tok/s | Single-worker risk |
| gpt-oss-120b | text | 1 | 3,491 | 3491.0 | 50.3 tok/s | Single-worker risk |
| qwen38-flash-next-125b-nvfp4 | text | 1 | 493 | 493.0 | 6.3 tok/s | Single-worker risk |
| qwen3-27b | text | 1 | 440 | 440.0 | 82.7 tok/s | Single-worker risk |
| FLUX.2 Klein 4B FP8 | image | 1 | 79 | 79.0 | 4.1s average | Single-worker risk |
| LTX-2.3 | video | 1 | 56 | 56.0 | 67.9s average | Single-worker risk |
| gpt-oss-20b | text | 1 | 37 | 37.0 | 157.4 tok/s | Single-worker risk |
| Krea 2 Turbo | image | 1 | 31 | 31.0 | 8.6s average | Single-worker risk |
| ace-step-v1.5-xl-turbo | audio | 1 | 25 | 25.0 | 14.1s average | Single-worker risk |
| z-image-turbo | image | 1 | 23 | 23.0 | 18.8s average | Single-worker risk |
| LTX Director 2.0 | video | 1 | 0 | 0.0 | No recent timing sample | Single-worker risk |
| LTX-2.3 Audio | video | 1 | 0 | 0.0 | No recent timing sample | Single-worker risk |
Media qualification progress
Needed right now
Help qualify the media worker.
The first managed media profile stays closed until accepted evidence covers every required GPU class. Each evidence set runs three local canaries and cannot enroll a worker or earn rewards.
Evidence needed
Midrange
20 GB to under 80 GB VRAM
0/1 accepted evidence sets
Evidence needed
Datacenter
80 GB or more VRAM
0/1 accepted evidence sets
Profile ace-step-v1.5-xl-turbo v0.2.4 · status updated Sep 4, 2026 UTC
Public operator proof
Check the rail before you commit a GPU
Verify that workers are visible and payouts are settling. The Grid does not turn hardware specs or historical demand into guaranteed earnings.
Payout scenario
- Settled pool, rolling 24h
- 4,120.9767 AIPG
- Observed periods
- 24 hourly settlements
Same-window scenario
41.2098 AIPG
If a worker had earned 1.0% of all accepted den during this exact window. This is arithmetic on settled history, not a payout forecast.
Den depends on completed work, model weighting, availability, competition, and successful settlement. Raw job count, GPU name, and token price are deliberately excluded. Last payment: Sep 6, 2026, 5:00 PM UTC.
Verify payouts on BaseLive worker check
Is your worker visible to the Grid?
Paste the exact worker name or worker ID printed after connection. The check reads the public online registry and sends no hardware details.
Start the worker first, then use the identity from its local dashboard or connection log.
How local checks and Grid connection work
Local decision
The browser does not decide what your GPU can run.
Private hardware check
The manager detects GPU, VRAM, driver, RAM, disk, and architecture locally. The Grid receives a capability tier and measured performance, not your complete hardware inventory.
Fail-closed profiles
Exact source, dependencies, model files, and recipes are verified before installation and again before serving. A capability stays unavailable until its local canary passes.
Operator flow
From your backend to Grid work
- 01
Connect locally
Start the runtime you already use, or install one separately. The worker detects its endpoint and available models locally.
- 02
Choose and test
Select what to advertise, set capacity limits, and run a real local generation before the worker connects to the Grid.
- 03
Verify and serve
Confirm the worker appears online, then serve compatible jobs. Completed accepted work contributes to the current payout split.