OneTriangle Inference

Model API

Inference service that prefills long inputs on a small model and decodes on a large one by transferring the KV cache between them, falling back to ordinary prefill when a model pair does not meet quality and latency gates, and also serves open-weight models such as DeepSeek V4 Flash, Llama 3.3 70B and Qwen3 variants on managed GPUs; priced per million input and output tokens, with a demo on request.

Find alternatives to OneTriangle Inference

TypeModel API

RoleStandalone tool

Bought forServe a model in production, Track and cut what we spend on AI models

AvailabilitySold

DeploymentSaaS

Intended forDeveloper, Enterprise

SecurityVendor-asserted

CertificationsNot recorded

Trains on customer dataNo

Zero data retentionUnknown

Data residencyNot recorded

Security featuresNot recorded

StatusActive

Sold withinSold on its own

PricingUsage-based — official pricing ↗

Pricing, Deployment, Security, Sold within — onetriangle.ai · 20 Oct 2026

What it does

Model inference
Model hosting

Features 4

KV cache transferPrefill runs on a small model and the cache moves to a large model for decode.
Quality gatesTransfers are gated on held-out quality and latency for each model pair, with ordinary prefill as fallback.
Serviced open-weight modelsRuns open-weight models on its own GPUs and serving stack.
Published per-token pricesLists Input, Cached input and Output prices per model.

onetriangle.ai

Sources

Pricing

Pricing page · 20 Oct 2026

Models page lists per-model Input, Cached input and Output prices, for example DeepSeek V4 Flash at Input $0.15 and Output $0.35; "All prices per M tokens".

Description

Official documentation · 20 Oct 2026

Home page: "Prefill runs on a small model, and decode on a large one."

What it does

Official documentation · 20 Oct 2026

Home page: "Prefill runs on a small model, and decode on a large one."

Constraints