FinHardBench - System Configuration: leaderboard
The design-space exploration experiment of FinHardBench (COLM 2026): the LLM chooses per-stage variants for a fixed 6-stage high-frequency trading pipeline, 432 configurations under a 5,000-LUT budget, and refines its choice over 24 rounds of timing feedback. Score is the best end-to-end latency reached, in nanoseconds, averaged over five seeds; lower is better and 107.5 ns is the optimum.
Metric: Mean best end-to-end latency in nanoseconds (lower is better; mean over 5 seeds of the lowest pipeline latency the model reaches in 24 feedback-driven rounds choosing among 432 configurations of a fixed 6-stage HFT FPGA pipeline under a 5,000-LUT budget; the optimum is 107.5 ns). Source: arxiv.org. Saturation forecast: Around May 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 107.5 | #161 |
| 2 | GPT-5.4 | 107.5 | #123 |
| 3 | MiniMax-M2.7 | 108.5 | #518 |
| 4 | DeepSeek V3.2 | 116.1 | #406 |
| 5 | Gemini 3.1 Flash Lite (Preview) | 121.4 | #418 |
| 6 | Mistral Large 3 | 129.4 | #817 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=finhardbench-system-configuration · How It Works · Data refreshed daily, snapshot 2026-11-13.