servershareOpen AI models on a
community-powered GPU network.
One OpenAI-compatible gateway for language, image and other open models — served by pools of independent GPU hosts, with every model request served in an isolated, ephemeral virtual machine. Build your own apps on it, run a model in a private pool of your own, or create in our studios.
Request API access Explore studios
Have an idle GPU or Mac? Become a host →
Invite-only test · Free test credits · No tracking cookies
Where do you want to go?
Use a model API
An OpenAI-compatible endpoint for open models. Get a key and send your first request.
Create your own pool
A private model pool only your keys can call, on machines you rent from hosts of the class you choose. A model too large for one host can be spread across several machines. In testing
Tokens on your own pool cost nothing: you pay rent for its machines, in test credits during the test.
Become a host
Connect your GPU or Apple Silicon Mac. Work runs in isolated virtual machines, never directly on your system.
Need a plain virtual machine instead? You can rent one from the host catalogue — see the rental guide.
You will be asked to sign in if you have not. No invitation yet?
Free test credits during the open test. Test credits only — no money is paid.
Change one line. Keep your code.
The ServerShare gateway speaks the OpenAI API. Point your existing SDK at our base URL and run open models on the network.
from openai import OpenAI
client = OpenAI(
base_url="https://servershare.io/v1",
api_key="YOUR_SERVERSHARE_KEY",
)
stream = client.chat.completions.create(
model="example-chat-model",
messages=[{"role": "user", "content": "Hello!"}],
stream=True,
)
for chunk in stream:
print(chunk.choices[0].delta.content or "", end="")| Endpoint | What it does | Status |
|---|---|---|
/v1/chat/completions |
Chat and text generation with streaming | In testing |
/v1/models |
Model catalogue with capabilities and limits | In testing |
/v1/images/generations |
Image generation from text, with style presets | In testing |
/v1/embeddings |
Text embeddings | Planned |
/v1/audio/transcriptions |
Speech to text | Planned |
/v1/jobs |
Long-running jobs, with a signed webhook when they finish. Video jobs are Planned. | In testing |
Streaming and usage counts In testing
Standard streaming responses with token usage on every request.
Seamless failover In testing
If a host drops mid-response, a streamed answer continues on another host in the same pool without repeating tokens. A request that fails before the first byte is retried on another host.
Usage you can trust In testing
Tokens are counted by the platform, not reported by hosts.
Custom models Planned
Serve your own fine-tuned adapters (LoRA).
Available models
The model catalogue is published to invited developers during the test.
| Model | Type | Context or max size | Pool status |
|---|
Model pools: many hosts, one endpoint.
Each model runs in a pool — a group of hosts with the same model loaded and ready. The gateway spreads requests across the pool, sends each one to the ready host with the lightest load and takes a host out of rotation when it stops answering.
Matched to hardware In testing
A pool machine can only be created on a host with enough GPU memory for the model.
Rated hosts first In testing
Machines below a pool's minimum rating get no requests, and new machines get less traffic. Among the rest, the shortest queue decides.
Always warm In testing
Pools keep models loaded, so there's no waiting for downloads.
| Host hardware | Typical workloads |
|---|---|
| GPUs with 12–16 GB | Smaller language models, image generation, embeddings |
| GPUs with 24–32 GB | Mid-size language models (quantized), image and video generation, custom model training |
| Multi-GPU hosts | Larger language models |
| Apple Silicon with large unified memory In development | Large mixture-of-experts language models |
Examples only. Actual models depend on the catalogue and on available hosts.
Holding a machine that serves a pool: the pools guide covers what is checked, how the rating is built and how served tokens are settled.
Private pools In testing
Create a pool only your keys can call: take a model from a platform pool, set the minimum host class, and put machines into it on hosts of that class or higher. You choose how the model is laid across the cards of a machine, and a model too large for one host can be spread across several machines that act as one copy.
Tokens served by your own pool cost nothing; you pay rent for its machines, in test credits during the test. Loading your own fine-tuned models is part of Custom models, which is Planned.
Studios — built on the same gateway.
Not a developer? Our studios give you the same network in a friendly browser app. They use exactly the same public API that developers get.
Design Studio In testing
Create images from text with style presets, repeat what worked from your history, download the results and rate or report each one. Editing, variations, upscaling and custom styles are Planned.
Chat Studio In testing
Chat with open language models, ask two of them the same question and save conversations — through the same gateway and keys as any app.
More studios Planned
Video, audio and specialised tools as the network grows.
Custom training is subject to our Acceptable Use Policy.
How it works
-
Request
Your app or a studio sends a request to the gateway.
-
Route
The gateway picks the pool serving the model and its ready host with the lightest load, skipping machines below the pool's minimum rating or trust tier In testing. The platform assigns each machine its tier when the machine starts, from the evidence it checks itself.
-
Run
The host serves the request inside an isolated, ephemeral virtual machine.
-
Return
Results stream back; usage is counted by the platform. In testing
-
Recover In testing
If a host fails, the request continues on another host.
Reliability rating In testing
Every pool machine is scored on uptime, successful requests, time to ready, speed and integrity checks. Machines below a pool's minimum get no requests; new machines start small.
In testing During the test, some requests may be served by third-party model providers when no suitable host is available. It is switched on per model by the platform, and every such request is recorded with the provider's name. See the Test Terms.
Why ServerShare
Drop-in compatible
Works with existing OpenAI SDKs and tools.
Full VMs, not containers
Every workload runs behind a hypervisor boundary.
Ephemeral hosts In testing
Hosts keep nothing after a job ends.
Open-source host agent Planned
Anyone will be able to review what runs on host machines.
Trust tiers
Different workloads need different protection. The gateway routes each request to a tier that fits it.
| Property | Community | Verified | Confidential |
|---|---|---|---|
| Status | In testing | In development | Planned |
| Hosts | Individual hardware owners | High-rated hosts with hardware attestation | Hosts with hardware-level memory protection |
| Protection | VM isolation, ephemeral hosts | + attestation, integrity checks | + hardware-protected memory |
| Used for | Public content, general creative work | Language-model requests, your own images, custom training | Sensitive data |
The platform assigns a tier to every machine when it starts, from evidence it checks itself: the host's TPM attestation, its rating and the integrity checks. A tier the host claims for itself is ignored, and a machine's tier can drop while it runs but never rise. No host can prove hardware memory protection yet, so no machine is Confidential.
Language-model prompts often contain private information, so they are routed to Verified hosts once that tier is available. Until then, please don't send confidential data during the test.
Security you can check, not just trust.
-
VM isolation In testing
-
Ephemeral hosts In testing
No user data remains on host machines after a job.
-
Restricted network access In development
A pool machine can be limited to the platform and approved model sources. Not every workload is limited yet.
-
Platform-side usage counting In testing
-
Encrypted storage In testing
-
Encrypted transport In testing
-
Integrity checks In testing
Control requests compare the machines of a pool with each other; a machine that disagrees is taken out of rotation.
-
Host attestation (TPM) In development
The host agent proves how the machine booted with its TPM chip. Renters are not shown it yet: it does not yet vouch for the agent itself.
-
Open-source host agent Planned
| Threat | Community | Verified | Confidential |
|---|---|---|---|
| Leftover data on host after a job | ✅ Protected | 🗓 Planned: protected | 🗓 Planned: protected |
| Workload escaping into the host system | ✅ Protected | 🗓 Planned: protected | 🗓 Planned: protected |
| Network eavesdropping | ✅ Encrypted | 🗓 Planned: encrypted | 🗓 Planned: encrypted |
| Host serving a different model than requested | ⚠ Spot-checked | 🗓 Planned: checked + attested | 🗓 Planned: checked + attested |
| Host inspecting a running request's memory | ❌ Not protected | 🗓 Planned: reduced risk | 🗓 Planned: hardware-protected |
Built for responsible use.
We do not allow:
- Generating or training on images of real people without their documented consent
- Any content that sexualizes minors
- Non-consensual intimate imagery or deceptive deepfakes
- Malware, spam, fraud or other abusive uses of the network
- Content that infringes others' rights
Requests are rate-limited and watched for abuse patterns. Image prompts and generated images are screened before anything is returned, and text prompts are checked by a classifier on our own node before they reach a pool In testing. Text screening is limited: it reads the start of a long prompt, does not check generated text, and most hits go to a person for review rather than being refused. Apps built on ServerShare must follow the same rules. Violations lead to removal and suspension.
Turn your idle hardware into part of the network.
ServerShare runs on machines owned by people like you. Install the agent, choose when your machine is available, and it joins model pools that fit its hardware — serving language models, image generation and more.
Why host with us
Your system stays yours
Workloads run inside a virtual machine with a hypervisor boundary — never directly on your operating system.
Nothing left behind In testing
Virtual machines are ephemeral. No user data stays on your disk; only verified public model files are cached.
Restricted network use In development
A pool machine can be limited to the endpoints it needs, so your connection can't be used for spam or attacks. Not every workload is limited yet.
You're in control In testing
Pause a pool machine at any time: it finishes the requests it has and takes no new ones, so you can stop it without cutting anyone off. Set availability hours, and outside them the machine leaves the pool the same way.
How it works
-
Message the bot
Tell us about your hardware in Telegram and get an invite.
-
Install the agent
The installer on Windows or the app on macOS. Linux is In development: machines without a GPU only, so a Linux host cannot serve pools yet.
-
Join a pool In testing
Create a pool machine for a model; the network checks that your GPU has enough memory for it.
-
Build your rating In testing
Uptime, successful requests and speed raise each pool machine's reliability rating; machines below a pool's minimum get no requests.
| Your hardware | Pools you can join (examples) |
|---|---|
| NVIDIA GPU, 12–16 GB | Smaller language models, image generation, embeddings |
| NVIDIA GPU, 24–32 GB | Mid-size language models, image and video generation, custom model training |
| Several NVIDIA GPUs in one machine | Larger language models |
| Apple Silicon Mac with large unified memory In development | Large language models |
Requirements
- NVIDIA GPU with 12 GB VRAM or more, or an Apple Silicon Mac
- Stable broadband, 100 Mbps+ upload recommended
- Virtualization enabled in BIOS/UEFI (NVIDIA hosts)
- Machine able to stay on while it serves a pool
| Platform | Status |
|---|---|
| Linux | In development |
| Windows | In testing |
| macOS (Apple Silicon) | In testing |
| Android | Planned |
Host questions
Will hosting slow down my computer?
Only while it's serving work. You can pause at any time.
What about electricity?
Hosting uses power while serving work. During the test there is no payout, so host only if you're comfortable with that cost.
Can I see what runs on my machine?
The panel shows which pool your machine is in, its state, the requests and tokens it served and what it earned in test credits. The agent source code will be published. Planned
Can I stop hosting?
Yes, at any time.
Tell the bot your GPU or Mac model, memory, operating system and upload speed. The host guide lists what a machine needs and what stays yours.
Roadmap
-
In testing
VM isolation, ephemeral hosts, Windows and macOS hosts, OpenAI-compatible gateway (chat, image generation, model catalogue and async jobs), API keys and rate limits, language and image model pools with pause and availability hours, private pools with a model spread across several machines, platform-side usage counting, test credits, host accelerator passport, telemetry and benchmarks, reliability ratings, failover mid-stream, integrity checks, screening of prompts and generated images, third-party model providers, Chat Studio, Design Studio, routing by the Community trust tier.
-
In development
Restricted guest network, Apple Silicon pools, Linux hosts (machines without a GPU so far), host attestation (TPM), Verified tier.
-
Planned
Embeddings and speech endpoints, video jobs, custom models (LoRA adapters), open-source agent, Confidential tier, paid plans and host payouts after the test.
Questions
What can I run on ServerShare?
Open models that fit the hardware in our pools — language, image and more. See the model catalogue for what's available right now.
Do I need a new SDK?
No. The gateway is OpenAI-compatible; change the base URL and API key.
Can hosts see my prompts?
Requests run in isolated, ephemeral virtual machines. A host with full control of its machine could in theory inspect a running request on Community hosts. Language-model requests will be routed to Verified hosts once that tier is available. Please don't send confidential data during the test.
Do you store my prompts?
The gateway does not keep the text of your prompts and answers — only that a request happened, which pool answered and how many tokens it cost. The studios keep your own history until you delete it. What is kept and for how long is in the Privacy Policy.
What happens if a host goes offline mid-response?
A streamed answer continues on another host in the same pool, without repeating tokens In testing. An answer that is not streamed is cut off, and your SDK sees an error you can retry. A host that fails before sending anything is replaced automatically.
Are requests always served by the ServerShare network?
Not always. When no suitable host is available, some requests may be served by third-party model providers In testing. The platform switches this on per model, and every such request is recorded with the provider's name. This is described in the Test Terms.
Is it free?
During the test, yes. Everything runs on test credits with no monetary value.
How do I become a host?
Message our Telegram bot using the Become a host button. See the host section for requirements.
Do hosts get paid?
Not during the test. Hosts receive test credits only.
How do I contact you?
Only via our Telegram bot for now.
Can I use outputs commercially?
It depends on each model's license, shown in the catalogue.
Join the test.
Tell the bot what you're building, which models you need and your expected volume.
I'm building with the API I want to use a studio I want to host
Opens our Telegram bot. Telegram's own privacy policy applies to messages sent there; see our Privacy Policy for how we handle them.
How to get access
- Message our bot on Telegram.
- Ask it for an invitation link.
- Open the link and set your own password.
Already have an account? Sign in
No payments and no payouts: usage is accounted for in test credits, which have no monetary value. This page has no sign-up form and collects nothing from you. What the platform stores once you are inside, and what you must not upload, is set out in the terms you accept when you sign in.