One cloud for compute, inference, and agents. GMI Cloud is an NVIDIA Reference Platform Cloud Partner running production AI workloads on H100, H200, and Blackwell — from a serverless token-billed API all the way down to bare-metal clusters with root access.
If it runs on the Hub, we can serve it. Below is everything you need to send your first request.
Start in 30 seconds
Our Inference Engine is OpenAI-compatible — point the SDK you already use at our base URL.
from openai import OpenAI
client = OpenAI(
base_url="https://api.gmi-serving.com/v1",
api_key="YOUR_GMI_API_KEY", # console.gmicloud.ai → API Keys
)
response = client.chat.completions.create(
model="deepseek-ai/DeepSeek-V4-Flash",
messages=[{"role": "user", "content": "Hello, what can you do?"}],
)
print(response.choices[0].message.content)
curl https://api.gmi-serving.com/v1/chat/completions \
-H "Authorization: Bearer $GMI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek-ai/DeepSeek-V4-Flash",
"messages": [{"role": "user", "content": "Hello, what can you do?"}]
}'
Tool calling, structured outputs and streaming all work exactly as they do against OpenAI — so the same two lines swap GMI Cloud into Claude Code, Codex, Cursor, OpenCode, Kilo, Factory, Dify and Hermes with no code changes.
Quickstart · LLM API reference · Rate limits · Get an API key
Guides — start from what you're trying to do
| If you want to… | Start here |
|---|---|
| LLMs & text | |
| Send your first chat completion | Quickstart · LLM API reference |
| See which text models are live, with the exact model ID for each | LLM model catalog |
| Try a model in the browser before writing any code | Playground |
| Understand rate limits and how to raise them | Rate limits |
| Choose between pay-per-token and reserved GPUs | Serverless vs Dedicated endpoints |
| Deploy your own weights behind an endpoint | My Models · Inference storage |
| Coding agents | |
| Point your coding tool at GMI Cloud | Claude Code · Codex · Cursor · OpenCode · Kilo · Factory |
| Build a multi-step agent or research workflow | Hermes · Dify deep research · OpenClaw |
| Publish or consume an agent | AgentBox · Register an agent · Long-running requests |
| Image, video, audio & 3D | |
| Generate or edit an image | Image models |
| Generate video from text, an image, or a reference | Video models · Video API · Video SDK |
| Synthesize speech, clone a voice, or generate music | Audio models |
| Chain models together without writing orchestration | GMI Studio · Studio tutorials |
| GPU compute | |
| Rent bare metal or a managed Kubernetes cluster | GPU Compute · Bare metal · Managed clusters |
| Set up networking and isolation | VPC & subnets · Firewalls · Elastic IP |
| Move your data over from S3 | Migration guides |
| Track spend and usage | Pricing · Usage |
All guides → · Full documentation → · llms.txt (point your agent at this)
What we run
⚡ Inference Engine — Serverless
Pre-configured, OpenAI-compatible endpoints. No infrastructure to manage, pay per token, scales to zero. Built-in batching and latency-aware scheduling.
Serverless docs →🎯 Inference Engine — Dedicated
Your models on dedicated GPUs. Full control over hardware, scaling, and isolation. No rate limits. For latency-sensitive production traffic.
Dedicated docs →🖥️ GPU Compute — Cluster Engine
Managed Kubernetes and bare metal on H200 and B200. Root access, custom stacks, RDMA-ready networking, VPC, firewalls, elastic IPs.
Cluster docs →🎨 GMI Studio
Visual workflow canvas for multi-step pipelines — chain image, video, audio, and LLM nodes without writing orchestration code.
Studio docs →🤖 GMI AgentBox
Marketplace for AI agents — search, run, register, and publish agents, with first-class handling for long-running requests.
AgentBox docs →📦 Model Library
Hundreds of ready-to-call models across every modality, each with a copy-paste quickstart. Test in the browser playground before you integrate.
Browse the library →
Model coverage
| Modality | What's live | Examples |
|---|---|---|
| Text | Chat, code, reasoning, agents — up to 1M context | DeepSeek V4 Pro & Flash · Kimi K3 · Qwen3.8-Max · GLM 5.1 · MiniMax M3 · Nemotron 3 Ultra 550B · MiMo V2.5 Pro |
| Image | Generation, editing, inpainting, upscaling, background removal | FLUX.2 · SeedDream 5 Pro · Qwen-Image · Z-Image Turbo · Wan 2.7 Image · Bria (Eraser, GenFill) |
| Video | Text-to-video, image-to-video, reference-to-video, lip sync, editing | Wan 2.7 · Kling V3 · SeedDance 2.5 · LTX-2 · Veo 3.1 · PixVerse V6 · SkyReels V4 · MiniMax H3 |
| Audio | TTS, realtime TTS, voice cloning, music | MiniMax TTS 2.6 & Voice Clone · MiniMax Music 2.5 · Realtime TTS 2 · Chatterbox |
| 3D | Asset generation | Hunyuan 3D Pro |
Full model library with per-model quickstarts →
Hardware
| GPU | Status |
|---|---|
| NVIDIA H100 | Available now |
| NVIDIA H200 | Available now |
| NVIDIA B200 / Blackwell | Available on Cluster Engine · pre-order for reserved capacity |
Deployed on NVIDIA Reference Platform Cloud Architecture as an NVIDIA Preferred Partner, with RDMA-ready interconnect and multi-tenant isolation for predictable tail latency. Current pricing →
Get started
Console · Docs · Quickstart · Playground · API reference · Pricing
💬 Join the GMI Cloud Discord — talk to our engineers, get help, and hear about new models first.
gmicloud.ai · X / @gmi_cloud · LinkedIn · Talk to us about dedicated capacity