
1. Technology overview
OneTriangle is a San Francisco start-up (Y Combinator batch Summer 2026, founded 2026 per YC) that optimises LLM inference. It addresses the cost of prefill on long inputs, where large-model rates apply before the model produces a first word.
The engine runs prefill on a small model and decode on a large one, transferring the KV cache between them through head-wise projection. The site reports results such as Llama 3.1 8B to 70B first-token time falling from 16.9s to 2.7s at 100k tokens. It integrates with dedicated endpoints or customer GPUs.
2. Highlighted features & specifications
- KV cache transfer: moves cache from a small to a large model.
- Split inference: prefill on small model, decode on large.
- Benchmarks: published latency and cost figures.
- Integration: works with dedicated endpoints or GPUs.
- Demo: interactive demo and research pages.
3. Potential application
A bank running a private LLM can cut latency on long document queries. A telecom can lower cost of customer-service assistants. A cloud provider can speed up hosted model endpoints.
4. Benefits
Reports first-token time 6.3x faster and prefill cost 84 percent lower in its benchmark. Integrates with dedicated endpoints or customer GPUs. Based on published cross-model KV cache transfer research.
5. Market trends & opportunities
Long-context LLM workloads raise inference cost, driving demand for serving optimisation. Vietnamese enterprises running self-hosted models on local GPUs look to cut serving cost.
6. Other information
Based in San Francisco, CA per YC; Y Combinator Summer 2026 batch. Benchmark figures are the company's own.
One platform connecting startups, corporations, research institutions, regions and investors











Playbook & Handbook · February 6, 2026
Winning Accelerator Playbook 2026 by BambuUP
RETAIL E-COMMERCE: AI AND REAL-TIME DATA ARE TRANSFORMING HOW RETAIL BUSINESSES OPERATE








































