Runcrate Inference Engine
Dedicated inference, sized to your traffic.
Llama, DeepSeek, Kimi, Qwen and GLM behind one OpenAI-compatible API. Pay per token first, then GPUs reserved for you at a committed rate.
Running a coding agent? Open models are free with your GitHub account. Big workload? Talk to an engineer below.
Endpoints
One API for every output. Text, images, video and sound.
/v1/chat/completions
Chat and reasoning, whole or streamed, with tool calls and JSON output.
79 text models
/v1/images/generations
Images from a prompt, or an edit of one you send.
24 image models
/v1/videos
Video from a prompt or a first frame, as a job you poll and download.
13 video models
/v1/audio/speech
Text read aloud, returned as audio.
mp3 · wav · opus · flac
/v1/audio/transcriptions
Audio in, its text out.
file upload, JSON back
/v1/embeddings
Vectors for search and retrieval.
JSON back
Change the base URL in an OpenAI SDK to https://api.runcrate.ai/v1 and existing calls keep working. GET /v1/models lists the catalog without a key.
Examples
Ask your agent. Or read the request it sends.
The same OpenAI-compatible request works on the pay-per-token API today and on dedicated capacity later.
Hand this to your agent
Add a Runcrate model to this project: run curl -fsSL"https://runcrate.ai/agents.md?intent=build"and follow it.
Your agent asks for an API key. Create a key
Pricing
The open frontier, at today's prices.
Model
Per 1M tokens, as billed
Llama 3.3 70B
Meta
$0.105 in
$0.336 out
DeepSeek V3.2
DeepSeek
$0.273 in
$0.399 out
DeepSeek R1 0528
DeepSeek
$0.525 in
$2.26 out
Kimi K2.6
Moonshot AI
$0.788 in
$3.68 out
Qwen3 235B A22B 2507
Alibaba
$0.0945 in
$0.578 out
GLM 4.7
Z.ai
$0.42 in
$1.84 out
gpt-oss 120B
OpenAI
$0.0389 in
$0.179 out
Text models bill per token, input and output priced apart. Images, video and speech bill per item or per second.
One dollar balance pays for every model and every GPU, with the same key.
Dedicated capacity is a monthly commitment at a discounted per-token rate. Usage above it bills at that same rate.
Options
Change only what you need. Everything else has a default.
- Pick the model
model - Default: required
- Any catalog id, like deepseek-ai/DeepSeek-V3.2. Change it and nothing else in your code moves.
- Stream the reply
stream - Default: false
- Tokens arrive as they are made, as server-sent events, the way OpenAI streams them.
- Cap the length
max_tokens - Default: model's
- The most tokens the reply may use, and so the most it can cost.
- Let it call your tools
tools - Default: none
- Functions the model may call, in OpenAI's format. The reply says which and with what.
- Get JSON back
response_format - Default: text
- Ask for a JSON object instead of prose, on models that support it.
- Set the randomness
temperature - Default: model's
- 0 for the same answer each time, higher for variety.
top_pworks too.
Under the hood
Arc, our serving stack. Built for production traffic.
Continuous batching tuned to real traffic, aggressive KV-cache reuse, speculative decoding, and quantization checked against quality. One request fans out across replicas sized to your load.
throughput
More tokens from each GPU you reserve, so fewer GPUs do the same job and the rate card shows it.
more tokens per GPU
capacity
GPUs reserved and sized to your traffic. No shared queue, no neighbor's burst in your latency, no rate-limit errors at peak.
yours alone
rate
A monthly commitment buys a discounted per-token rate, agreed up front. No capacity units to translate.
committed spend
Compared with shared APIs
Dedicated, not pooled. Here is what changes.
| Compared on | Typical shared API | Runcrate Inference Engine |
|---|---|---|
| Capacity | A shared serverless pool | Dedicated, sized to your workload |
| Latency at peak | Moves with other tenants' traffic | Held by capacity reserved for you |
| Pricing | Per-token list rates | A committed-spend rate |
| Rate limits | Token ceilings, errors on burst | Sized for your bursts |
| Custom models | Limited, or extra | Bring your weights, we serve them |
| Support | A shared queue | A named contact and an on-call engineer |
Talk to an engineer
Size your deployment.
Tell us your model, your spend and your traffic. Within 24 hours you get a rate card and a 7-day pilot plan that runs beside your current provider, with no commitment.
FAQ