OneTriangle
OneTriangle is a YC Summer 2026 (S26) company. Based in San Francisco, CA, USA. Team of 4.
About OneTriangle
Summary
OneTriangle builds an inference engine that cuts the cost and latency of running large language models by splitting the prefill and decode steps across model sizes. During prefill, a small model processes the input and its KV cache tensors are reprojected into the large model's attention space, so the large model begins decoding against a populated cache it never had to compute. This replaces the standard approach where the full large model handles every token of a long prompt at large-model rates.
Business model
A managed hosting service for open-weight models, accessed by changing one line of config.
Go-to-market
Teams running open-weight models such as Llama, Qwen, Mistral, Gemma, and DeepSeek are the target customers; they onboard via a demo booking flow and a self-serve config change, with OneTriangle handling GPU provisioning and serving infrastructure.
From the Y Combinator listing
The cheapest, fastest lightweight inference.
Inference has surpassed training as AI’s largest compute cost, accelerated by increasing demand for AI and agents. By using proprietary KV cache transfer tech, OneTriangle has cut inference costs by 20% and time by 40% compared to present standards.
Facts
- B2B / B2B -> Infrastructure
- San Francisco, CA, USA
- 4
- Active
- Early
Founders (2)
Score every batch against your own thesis
GPAgent's sourcing agent reads every company in a batch, scores each one against your fund's thesis in a private workspace, and lines up the founders worth meeting before Demo Day. You decide who to contact. Your scores, notes, and pipeline never appear on these pages.
How Accelerator intelligence works