SERVICE 01 · SERVERLESS INFERENCE

One API. Compatible with OpenAI and with Anthropic. Served from Europe.

Change the base URL and you are already running open models on our infrastructure. Nothing to deploy, no GPUs to manage, no commitment. You pay per million tokens.

Get your API key See models and prices

Reply within 48 hours.

The service

Four pieces, one single API.

api.nextbit256.com

from openai import OpenAI

client = OpenAI(
    api_key="nb-...",
    base_url="https://api.nextbit256.com/v1",  # the only change
)

SERVED FROM THE EU·Qwen3.5-35B-A3B·TTFT 210 ms·0.00% tool-call errors

Dual compatibility

Change the base URL. Your code stays put.

We expose the API in the two formats your stack already uses: OpenAI’s and Anthropic’s. If you build with Claude Code, Cursor, Cline or any agent that speaks either protocol, you change the base URL and the key and it keeps working.

< €20/mo25 people · 60 queries a day · mid-sized open model

Pay per token

No fixed fees. No minimums.

You pay per million input and output tokens, for what you use and nothing else. If you have an AWS spend commitment, you can buy via Private Offer on AWS Marketplace.

Our own infrastructure in Spain

Where it runs

Spain by default.

If any model in the catalogue were served from outside the European Union, it is flagged on its sheet and in the model table — before you integrate it, not after.

Load balancingAutoscalingQueuesFault toleranceObservabilityCompliance

What we run for you

Hardware, balancing and queues. You send an HTTP request.

The inference stack, the autoscaling and the fault tolerance are on us. You deploy nothing and manage not a single GPU.

01Create your account
02Change the base_url
03Runs in Spain
04Answers in ms

Who it is not for: if your load is high and sustained (an agent in production, a contact centre, a document pipeline), paying per token ends up costing more than Dedicated Inference. We tell you on the first call, even when it suits us less.

Catalogue

Available models, with their price per token.

They all speak the same API. Each model’s provenance is stated on its card, not in the small print.

Texto

Qwen3.5-35B-A3B

Qwen · 35B (A3B)

0,088 $/M entrada · 0,348 $/M salida

Texto

Gemma 4 26B A4B

Gemma · 26B (A4B)

0,100 $/M entrada · 0,400 $/M salida

Texto

ALIA-40B

ALIA · 40B

0,120 $/M entrada · 0,480 $/M salida

Texto

Kimi K2.5

Kimi · MoE

0,140 $/M entrada · 0,560 $/M salida

Multi

Texto

DeepSeek V4 Flash

DeepSeek · MoE

0,140 $/M entrada · 0,280 $/M salida

Texto

Llama 3.1 70B

Llama · 70B

0,150 $/M entrada · 0,600 $/M salida

See and compare all models

FAQ

The objections, answered.

Is it really compatible with the OpenAI API?

Yes: the same request and response format. You change base_url and api_key, and the rest of your code stays as it is.

Is my data used for training?

No, never. Neither prompts nor responses train any model, ours or anyone else’s.

How long do you keep requests?

90 days of operational retention with automatic deletion. On Dedicated Inference zero retention can be agreed.

Can I move to dedicated without rewriting code?

Yes. Same API and same layer: going from serverless to dedicated is a deployment decision, not a migration.

What rate limits are there?

Standard per-key limits meant for production. If your case exceeds them, they are raised during sizing.

Start now

Generate your API key and change one line of configuration.

No contracts, no minimums, not a single GPU to manage. Try it in minutes with your current stack.

No GPU managementPay per tokenServed from Spain

Get your API key Talk to us