HybridAgent Estimates Up to 72.9% Less Decoding Compute in Tests

HybridAgent used an estimated 62.5% to 72.9% less decoding compute than matched text agents in tests, while also delivering faster inference on three Qwen3 model sizes. The results suggest a possible efficiency gain for AI agents, though the cost figures are modeled rather than measured in deployed services.

The system was evaluated on 11 benchmarks using Qwen3-Base models at 4B, 8B and 14B parameter scales. Across the tests, HybridAgent showed an average performance improvement of 10.4%, used 35.5% to 54.7% fewer tokens and ran inference 1.8 to 2.6 times faster.

Estimated compute per question declined from 51.2 to 13.9 teraflops for the 4B model, from 62.5 to 17.2 teraflops for the 8B model, and from 66.9 to 25.1 teraflops for the 14B model.

HybridAgent combines latent reasoning, which allows some internal steps to occur without being decoded as ordinary text, with text generation for tool calls and final answers. Its HybridLink component moves between the two modes, while continuous working memory carries mixed context across turns.

The paper’s cost model puts the expense of one million questions at $98 to $176 with HybridAgent, compared with $359 to $470 for text-based systems. The estimates assume $2.50 per H100-hour and 10% utilization of peak BF16 throughput. They exclude prompt prefill, context-dependent attention and external tool execution, so they do not represent measured bills for production services.

The paper was submitted to arXiv on Oct. 13, 2026. Its findings cover the tested Qwen3 models; the results do not establish savings for larger models or deployed workloads.