Tencent Hunyuan's Accelerate Bench tests 16 coding agents on 32 CPU optimization tasks

Tencent-Hunyuan · hf · 2026-11-17

Tencent's Hunyuan team introduces Accelerate Bench, a benchmark of 32 CPU optimization tasks adapted from scientific applications and supercomputing contests, where agents choose where and how to optimize a working program under correctness and timing constraints. Evaluating 16 model-harness pairs, Score@1 ranges from 25.5% to 67.4%. Most outputs are correct, but speedups diverge: the top two agents achieve 2x or better on 22+ of 32 tasks versus 8–17 for the rest. Controlled replays show identical code runs 0.29x–2.12x depending only on build/runtime settings, and one change's payoff swings from 1.01x to 6.4x depending on another.

Original post →

More from coding & agent

coding & agent channel →