Fast, honest inference on open-weight models.
BlazingNodes runs open-weight models on our own GPU nodes and sells tokens through OpenRouter, with an OpenAI-compatible API. We never store prompts or completions, and every model states its quantization.
Our first two models are being tested on our first GPU node. Prices, served context and measured speeds go on this site before launch.
What you can count on
Fast where it counts
Agents, coding tools and chats resend long prompts. Prefix caching stays on, so repeated context returns quickly and is billed at the cached-input rate.
Busy means 429, not a timeout
We cap load before the GPU saturates. A full node answers with 429 and a Retry-After header before any work starts, never a dropped stream.
Zero data retention
Prompts and completions are processed in memory and never written to disk, logged or used for training. We keep billing metadata only.
Quantization, disclosed
Every model states its precision (FP8, BF16), the context we serve and its output limit. If one of them changes, the model page changes first.
Per token, no minimums
Three rates per model: input, cached input and output, in USD per million tokens. Failed and rejected requests are not billed.
The API you already use
OpenAI-compatible chat completions with streaming and token usage on every response. Change the base URL and the model id; keep your code.
First models
Both are open-weight under Apache 2.0 and will be served in FP8.
Call it like any OpenAI model
Through OpenRouter, pin requests to BlazingNodes with the provider order. Direct access athttps://api.blazingnodes.com/v1 follows the same format.
curl https://openrouter.ai/api/v1/chat/completions \
-H "Authorization: Bearer $OPENROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "google/gemma-4-26b-a4b-it",
"messages": [{"role": "user", "content": "Summarize this ticket in one line."}],
"provider": {"order": ["blazingnodes"], "allow_fallbacks": false}
}'The provider name blazingnodes applies once our OpenRouter listing is approved.