Turn GPUs into tokens.
Nimbo Turbine serves leading open models behind one OpenAI-compatible API. Pay per token, move to a dedicated endpoint when traffic is steady, and fine-tune without managing a GPU. No MLOps team required.
Change the base URL. Keep your code.
The API follows the OpenAI Chat Completions and Embeddings formats, so the SDKs, frameworks and agents you use today work unchanged.
- Streaming, tool calling and JSON modeSupported on models that support them.
- Data stays out of trainingPrompts and outputs are never used to train models. Zero-retention is available per key.
- Region pinningChoose where requests are served: Nimbo regions or your own hardware through Fleet.
from openai import OpenAI client = OpenAI( base_url="https://api.nimbokinetics.com/v1", api_key=os.environ["NIMBO_API_KEY"], ) resp = client.chat.completions.create( model="qwen3-235b-a22b", messages=[{"role": "user", "content": "Summarise this contract"}], )
curl https://api.nimbokinetics.com/v1/chat/completions \ -H "Authorization: Bearer $NIMBO_API_KEY" \ -H "Content-Type: application/json" \ -d '{"model":"gpt-oss-120b","messages":[{"role":"user","content":"Hello"}]}'
import OpenAI from "openai"; const client = new OpenAI({ baseURL: "https://api.nimbokinetics.com/v1", apiKey: process.env.NIMBO_API_KEY, });
Speech in, speech out, on the same API
Transcribe calls and meetings, then read answers back in a natural voice. Voice runs on the same keys, regions and bill as text models, and powers the Nimbo Secure copilot.
| Model | Task | Price |
|---|---|---|
| Whisper Large v3 | Speech to text, 99 languages | $0.003 / audio minute |
| Whisper Large v3 Turbo | Low-latency transcription | $0.0015 / audio minute |
| Kokoro-82M TTS | Text to speech, streaming | $15 / 1M characters |
Pay for tokens until a dedicated endpoint is cheaper
Per token
Shared, autoscaled endpoints for popular open models. Best for prototypes and spiky traffic.
from $0.14 / 1M input tokens
Per GPU-hour
Your model on reserved GPUs with vLLM or TensorRT-LLM. Predictable latency, no noisy neighbours, autoscaling within your reservation.
from $2.99 / GPU-hour
On your hardware
The same API served from your own GPUs through Nimbo Fleet, with overflow to Nimbo. One rate card, one set of keys.
platform fee per GPU, quoted
Rule of thumb: a dedicated endpoint pays off once sustained traffic keeps a GPU above roughly 40 to 50% busy. We will run that math with your traffic.
Bring data, get an endpoint
Upload a dataset, choose LoRA or full fine-tuning, and deploy the result to a dedicated endpoint in the same account. Your weights are yours to export.
LoRA and QLoRA
Fast, low-cost adaptation for most enterprise tasks.
Full fine-tuning
Multi-node runs on InfiniBand clusters for larger changes.
Evaluations
Compare base and tuned models on your own test set before you ship.
Exportable weights
Download adapters and merged weights. No lock-in.
Your first million tokens, today.
Create a key, change one URL, and see your workload running on Nimbo before you talk to anyone.