Attached document
One prompt, three inference paths
Inference demo Document Q&A
What can I help with?
Watch the prompt type, send, and run through all three paths automatically.
Three responses
Inference Pathways
Small is not decision-grade. Large is accurate but slow. OneTriangle delivers the large-model answer faster and for less.
Path 01
Small model
Qwen3 4BAnswer · incomplete
Path 02
Large model
Qwen3 14BAnswer · decision-grade
OneTriangle
Path 03
OneTriangle inference
Prefill small · decode large Small prefillKV transferLarge decode
Answer · same as large model
Run it yourself, live
Ask your own text or spoken question and compare our path against the native small and large models with measured TTFT, throughput, and outputs.
Cross-model KV cache transfer
Method, mapper architecture, and the published Qwen3 4B → 14B long-context benchmark.