Rent GPUs.
Run any model.
Ship AI.
ex5.ai is a GPU marketplace, a unified model API, a fine-tuning pipeline, a chat workspace and a coding CLI — one credit balance powers all of it. Top up with card.
npm i -g @ex5/clicurl api.ex5.ai/v1/chat/completions -d '{"model": "ex1-72b", ...}'Bare-metal GPUs from verified hosts, billed per second — H200 down to RTX 4090.
One OpenAI-compatible endpoint for EX-1 and 50+ open models. Pay per token.
Upload a dataset, tune EX-1 or any open model, get a private endpoint — one command.
A fast web workspace for every model, and an agentic CLI for your terminal.
Live GPU market
Real-time on-demand and interruptible pricing across regions. Rent in seconds, pay per second, cancel anytime.
prices refresh every 60s · all hosts verified & benchmarked · datacenter tier only
Every model, one API
Swap the base URL and keep your code. Our flagship EX-1 plus the best open models, priced per million tokens and billed from the same balance as your GPUs.
curl https://api.ex5.ai/v1/chat/completions \
-H "Authorization: Bearer $EX5_KEY" \
-d '{ "model": "ex1-72b", "messages": [...] }'Specialized, not general-purpose
Task-tuned models we train in-house — smaller, faster and cheaper than a frontier model on their own turf. Every one has a playground in chat.ex5.ai and a stable API id.
Your data in.
Your endpoint out.
No cluster setup, no training scripts. Fine-tune EX-1 or any open model on marketplace GPUs at spot prices, and we deploy the result as a private OpenAI-compatible endpoint the moment training finishes.
- 01Upload a JSONLDrag a dataset into the console or push it from the CLI — we validate, split and tokenize it.
- 02One command to trainLoRA or full fine-tune with live loss curves. Cost is estimated up front and billed per second.
- 03Auto-deploy on finishA private, autoscaling endpoint on api.ex5.ai — or download the weights and take them anywhere.
$ ex5 tune start --base ex1-72b --data ./support.jsonlA workspace for every model
Talk to EX-1, your fine-tunes or any catalog model — files, code and web search included. Metered per token, billed from the same balance.
__global__ void fused_topk(float* logits, int k) {
__shared__ float buf[1024];
// warp-level bitonic sort, no global sync
...Your terminal is the console
Everything on ex5.ai is scriptable: rent GPUs, stream logs, launch fine-tunes, deploy endpoints. And ex5 code is an agentic coding session in your repo, powered by EX-Code 32B.
- ✓ ex5 gpu — rent & ssh in seconds
- ✓ ex5 tune — managed fine-tuning
- ✓ ex5 deploy — models, containers, endpoints
- ✓ ex5 code — agentic coding in your repo
$ ex5 login ✔ authenticated as dev@acme.co · balance $48.20 $ ex5 gpu rent --type rtx4090 --count 2 --max-price 0.40 ✔ 2× RTX 4090 @ $0.34/hr · tr-izmir-2 · ssh ready in 41s $ ex5 tune start --base ex1-72b --data ./support.jsonl ⠸ ft-2841 running · epoch 2/3 · loss 0.412 $ ex5 deploy ft-2841 --name acme-support ✔ live → api.ex5.ai/v1 · "model": "acme-support" $ ex5 code "fix the race condition in worker.rs" agent reading worker.rs · 3 files changed · tests passing ✔ ▊
Our research is public
Weights, serving stack and tooling — free to download, self-host and modify. Or skip the setup and deploy any repo to a marketplace GPU in one click.
EX-1 serving recipes and eval harness.
$ git clone github.com/ex5-ai/ex1Deploy on GPU →The ex5 command line — gpu, tune, deploy, code.
$ git clone github.com/ex5-ai/ex5-cliDeploy on GPU →The template catalog behind one-click deploys.
$ git clone github.com/ex5-ai/templatesDeploy on GPU →One balance, three steps
Card or wire — prepaid credits in seconds, $5 minimum.
Raw GPUs, the model API, a fine-tune job or the chat workspace — mix freely, same balance.
Per-second GPU billing, per-token inference, one invoice. Scale to zero when you're done.
.aiStart building on ex5.ai
$5 free credits on signup. A 4090 for 15 hours, or about 9M tokens of EX-1 — your call.