Anthropic’s original performance take-home, now a community leaderboard. Hand-optimize a kernel for a simulated wide-issue VLIW machine and chase the lowest cycle count.
Your perf_takehome.py builds the instruction stream; the judge scores the worst cycle count across nine seeded validation runs — lower is better. Model reference runs on the board show where the frontier sits. Check out the GitHub repo to get started.
Community project — not affiliated with or endorsed by Anthropic.
Ranked by clock cycles (lower is better) — each account’s best score counts
| Rank | Author | Cycles | Attempts |
|---|---|---|---|
| #1 | @dougallj | 950 | 34 |
| #2 | @junlin_ai | 957 | 14 |
| #3 | @SaifAlHarthi | 958 | 54 |
| #4 | @NaregAM9 | 966 | 12 |
| #5 | @glenmtaggart | 970 | 32 |
| #6 | @jamespayor | 972 | 55 |
| #7 | @tlz173896 | 978 | 3 |
| #8 | @josusanmartin | 979 | 71 |
| #9 | @XYLi_Bruce | 979 | 2 |
| #10 | @Paul1365972 | 984 | 25 |
Model rows are reference runs shown for orientation only — single data points under specific harness conditions. Replication attempts may yield different results.