MODEL 01
GPT-OSS 20B
A compact model contributing an additional perspective to every query.
Candidate response MULTI-MODEL INTELLIGENCE, ONE APIMany perspectives.
Many perspectives.
One evaluated insight.
Build with answers that have been considered from more than one angle. Three models, coordinated orchestration, and a dedicated evaluation layer. All through one API.
Python & TypeScript Your code, your context One final answer
INFERENCE ENGINEARCHITECTURE
A considered answer.
A simpler integration.
Your application sends the question.
Inference coordinates the rest.
03Candidate models
01Evaluation layer
01Unified API
THREE PERSPECTIVES. ONE WORKFLOW.GPT-OSS 20BGPT-OSS 120BNemotron 3 Ultra
FROM QUESTION TO ANSWER
Intelligence is a process.
A structured workflow turns your context into a considered response.
Your team integrates once. Inference handles the coordination.
The current query flowMANAGED ORCHESTRATION
- 01
Query & context
Start with a question. Add code, a language, or reference context.
- 02
Orchestration
One request calls three candidate models in parallel.
- 03
Model responses
Three distinct perspectives address the same question and context.
- 04
Evaluation
Multiple available answers are evaluated; a single available answer is returned directly.
- 05
Final answer
The strongest acceptable response is selected and delivered to your application.
THE MODEL ENSEMBLEThree models.
Three models.
A broader perspective.
Each model responds to the same request and supporting context. The evaluation layer reviews the candidates, so your application receives a single selected answer.
MODEL 02
GPT-OSS 120B
A larger model contributing a separate response to the same context.
Candidate responseMODEL 03
Nemotron 3 Ultra
An additional model broadening the set of candidate responses.
Candidate responseBUILT AROUND ANSWER QUALITYOrchestration coordinates.
Orchestration coordinates.
Evaluation decides.
Generation and selection have distinct roles. A dedicated review step
assesses the candidate answers before one reaches your application.
EVALUATION LAYER
From candidates to one answer.
When multiple candidates are available, they are reviewed against the original request. A single available candidate is returned directly without evaluation.
- 01Validate
Exclude failed, empty, or mismatched responses.
- 02Assess
Review correctness, relevance, completeness, and requested format.
- 03Select
Choose the strongest acceptable answer without merging or rewriting it.
Multiple available answers are evaluated; a single available answer is returned directly.
DESIGNED FOR DEVELOPERSLess orchestration code.
Less orchestration code.
More product.
Keep your application stack. Send a question and optional context through a straightforward HTTP request.
Your application. Your access.
Create a platform key, keep it in your server environment, and authenticate your requests. Manage up to 3 keys from one console.
Manage API keys Account-scoped access Create and revoke keys 10 requests / minute · 100 requests / day
02 / INTEGRATE
One request. A complete workflow.
Start with copyable Python and TypeScript examples.
backend.tsHTTP API
// Run in your backend with a platform API key.
const response = await fetch(baseUrl + "/api/v1/queries", {
method: "POST",
headers: {
Authorization: "Bearer " + apiKey,
"Content-Type": "application/json",
},
body: JSON.stringify({ query, language, code, responseType: "text" }),
});YOUR DEVELOPER WORKSPACE
Everything you need to get connected.
Manage application access, explore the architecture, and find your integration examples.
Your backend. One HTTP API.
Python and TypeScript integration. The platform handles the models.
YOUR BACKEND SENDSPOST /api/v1/queries
Authorization: Bearer <API_KEY>{
"query": "Identify the failure modes in this implementation.",
"language": "javascript",
"code": "function first(items) { return items[0].name; }",
"url": "https://example.com/docs",
"responseType": "text"
}Evaluated response
One selected answer, reviewed for correctness, relevance, and completeness.
Three candidate answers · One final response01Query & context
02Orchestration
03Model responses
04Evaluation
05Final answer
A focused console for your account, API keys, and integration.
A LITTLE MORE CONTEXTGood questions.
Good questions.
Clear answers.
What to know before your first request.
Do I need to choose a model?
No. The orchestration layer calls three models in parallel for every query. Multiple available answers go through evaluation; a single available answer is returned directly.
How does evaluation work?
When multiple candidates are available, evaluation checks correctness, relevance, completeness, and format against your original request. Failed calls and null answers are excluded first. A single available answer bypasses evaluation. Text mode returns answer text; JSON mode parses it into structured data.
Can I include source code or documentation?
Yes. Add source code and its language as optional fields, or include documentation text in your query. You can also supply a reference URL. Its contents are not automatically fetched, so include any text the models need.
What are the request limits?
Each account has an allowance of 10 requests per minute and 100 requests per day, shared across its keys and authenticated API routes. If you reach a limit, the API returns a retry time.
Where do I start?
Open the console, sign in, and create an API key. The integration docs include Python and TypeScript examples for sending your first request from your backend. Request history and usage reporting are not available yet.
One integration.
More perspective.
Bring your question. Let Inference coordinate the answers.