OneTriangle

Experiments, graphs, measured results, and engineering notes from OneTriangle's model infrastructure research.

Join OneTriangle to build portable context and faster, cheaper AI inference.

OneTriangle works on optimizing prefilling computation, providing companies with the cheapest and fastest inference.

OneTriangle optimizes multi-model inference by transferring KV cache between models, cutting redundant prefill cost.

OneTriangle describes a method in which an 8B Ultravox model processes audio once and transfers its audio KV cache to a frozen 1B decoder. A mapper plus rank-16 residual translates the cache, and the system nearly doubles sealed-test accuracy while running at 133–150 tokens per second.

Sep 2026 · OneTriangleNews article

OneTriangle reports that DFlash generates the same tokens in less than half the time on Qwen3-8B. The system predicts 16-token blocks in parallel, achieving a 2.154× geometric-mean speedup and matching native token IDs in all 12 confirmation runs; the results appear in research note 013.

Sep 2026 · OneTriangleNews article

OneTriangle reports that its speculative KV-cache transfer method uses Qwen3-4B to prefill and draft tokens for a repaired Qwen3-32B verifier. In research note 012, the setup matches native Qwen3-32B task accuracy, delivers 19% higher throughput, and reduces long-context latency by 15.1%.

Aug 2026 · OneTriangleNews article

OneTriangle reports that Ultravox 70B builds a richer audio KV cache and transfers it to an 8B model, which generates at near-native 8B speed. The approach reaches 81.7% accuracy at 59.6 tokens per second and recovers 12.7 points over the native 8B model.

Aug 2026 · OneTriangleNews article