AI infrastructure
made for developers

Call open-weight models or push your own, scale to zero, run batch jobs and sandboxes, from the Python code you already write.

€5 in free credits for every new customer.

Get Started
gladiaOVHcloudQwant

Get it done, your way

Watch a request go through.

The OpenAI SDK call you already write, pointed at Blade. Each step below plays as it runs.

  • Add chat, reasoning and tool calling to your product through an OpenAI-compatible API.

  • Transcribe recordings or turn text into speech. Add voice features through the same API.

  • Extract text from documents with OCR, understand images.

  • Generate new visuals, billed per image.

  • Get answers with cited sources. Available on enabled accounts and supported models.

Illustrative films of the product.

Start with a model.
Or bring your own code.

Models

Available now

Open-weight models.
OpenAI-compatible API.

Run LLM inference with the OpenAI SDK you already use. Choose a model, set your Blade API key and base URL, and start building.

Add web search with Flash by Staan to get answers with cited sources. Available on enabled accounts and supported models.

Explore Models Read the API docs
python
import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.models.blade.sh/v1",
    api_key=os.environ["BLADE_API_KEY"],
)

response = client.chat.completions.create(
    model=os.environ["BLADE_MODEL"],
    messages=[{"role": "user",
               "content": "Hello, Blade."}],
)

AI Fabric

Opening gradually

Deploy your models.
Run Python on GPUs.

Serve your own models, process GPU batch jobs or run agent code. Start with a Python function: Blade builds the image, schedules the GPU and scales to zero. Join the waitlist to confirm your runtime and GPU requirements with the team.

Join the AI Fabric waitlist

Your Python workload. Ready for onboarding.

python
def transcribe(url: str):
    return model.transcribe(url)

Illustrative function. SDK setup and GPU access are provided during onboarding.

Work with many keys at once.

Give each key a job. They run in parallel, sleep when nobody calls, and wait for a quieter hour when they can.

One key, one job

Name your keys. The console shows what each one did and what it cost, request by request.

  • support-agent0€0.00000
  • transcriber0€0.00000
  • ocr-batch0€0.00000
  • image-studio0€0.00000

Illustrative requests, priced at the catalogue’s From prices.

The quieter hour, for flexible work

Dynamic rates follow GPU demand. Batch work that can wait runs at a quieter hour: same models, same results.

00:00quiet hour24:00

Illustration: the shape of demand, not a rate schedule.

Scale to zero

No traffic, no GPU. With AI Fabric, the next request wakes it up.

GPU0 %

Keys hand work to each other

Batch OCR reads the invoices, Support agent answers the customer. You approve the result, not every step.

Why Blade

Same models.
A bill you can read.

Usual inference APIs

  • One price, day and night
  • A price you have to look up
  • A plan before your first call
  • Idle GPUs on the bill
  • Your requests, somewhere

Blade

  • Dynamic rates that follow GPU demand
  • A From price published for every model
  • No subscription, no minimum spend
  • Scale to zero with AI Fabric
  • Shadow-operated GPUs in Europe

Same models.
Cheaper hours.

Models dynamic rates change with the hour and GPU demand. Shift batch OCR, embeddings or evals to quieter hours and pay less.

Compare model rates

From prices. Clear units.

Chat & codeEUR / 1M input tokens
From €0.0183
Vision & OCREUR / 1M input tokens
From €0.0232
Text embeddingsEUR / 1M input tokens
From €0.00774
Speech & audioEUR / audio minute
From €0.000155
Image generationEUR / image
From €0.00132

Lowest From price per category. Input rates shown for token-billed models.

Everything your key does, with its price next to it.

Requests, tokens and what each one cost, key by key, in the console.

  • 16 models
  • 328 ms fastest first token
  • From €0.0183 lowest, per 1M input tokens
  • €0 subscription

Why the rate moves. Dynamic rates follow GPU demand. The multiplier in force when you submit a request is the one used for it, and recorded usage and charges are in the console.

Illustrative usage: example requests priced at the catalogue’s From prices
TimeModelKeyUsageCost
13:02:11qwen3.6-35b-a3b support-agent2 140 tokens€0.00033
13:02:09whisper-large-v3-turbo transcriber4.0 min€0.00062
13:01:58qwen3-embedding-8b support-agent18 400 tokens€0.00014
13:01:40deepseek-ocr-2 ocr-batch7 110 tokens€0.00016
13:01:12flux-2-klein image-studio1 image€0.00132
Illustrative console view. Example requests, priced at the From prices of the catalogue (2026-10-22).

Put models to work.

A weather question calls get_weather and returns a weather response.LLM inference

Build assistants that act.

Add chat, reasoning and tool calling to your product through an OpenAI-compatible API.

Explore chat & tools
A microphone and waveform turn speech into a written transcript.Speech to text & text to speech

Let your app listen and speak.

Transcribe recordings or turn text into speech. Add voice features through the same API.

Explore speech & audio
Computer vision identifies a bicycle and a plant in an image.Vision, OCR & image generation

Read images. Create visuals.

Extract text from documents with OCR, understand images, or generate new visuals.

Explore vision & OCR
Agent code runs inside an isolated environment and returns a result.Agent sandboxes

Let your agents run code.

Prepare isolated code execution for AI agents, from Python jobs to GPU tasks. AI Fabric access is opening gradually.

Explore agent sandboxes

Hosted in Europe.
Operated by Shadow.

Know where your workloads run.

Model inference on Shadow-operated GPUs runs in Europe. Optional web search uses Staan. Check the data-handling details for model requests and external tools.

FAQs

Is it really the OpenAI API?

Models is OpenAI-compatible. Point the OpenAI SDK you already use at Blade’s base URL, set your Blade key, and keep your code.

Why does the rate change during the day?

Dynamic rates follow GPU demand. The multiplier in force when you submit a request is the one used for it, and each model publishes a From price: the lowest its rate can reach. How hourly pricing works.

Where do my requests run?

Model inference runs on Shadow-operated GPUs in Europe. Optional web search uses Staan. Read the data-handling details.

Is there a subscription or a minimum spend?

No subscription and no minimum spend. You pay for what you use, at the rate in force when you submit each request, and track it in the console.

Can I run my own code, not just models?

AI Fabric runs your Python on GPUs, scales to zero and handles batch jobs. It is opening gradually: join the waitlist.

Contact the team

Loading the form…

Dev Run

Loading the game…