LongCat 2.0

Meituan12 audited · none flagged

One row per published session deck. Score bars are peak fraction of the board roofline (Hard / CUDA) or best speedup vs the torch baseline (Mega). Audit chips come from the human/subagent reward-hack review of every published cell.

hard

longcat-claude/LongCat-2.05/60.4512 audited

RTX PRO 6000

FP8 GEMMpass
32.9%clean
session 1h 43m
solutiontrace
KDA CUTLASSpass
0.1%clean
session 2h 23m
solutiontrace
Paged Attentionpass
31.9%clean
session 3h 18m
solutiontrace
TopK Bitonicfail
fail
session 3h 43m
solutiontrace
Sonic MoEpass
7.1%clean
session 4h 30m
solutiontrace
W4A16 GEMMpass
10.1%clean
session 9h 20m
solutiontrace

H100 PCIe

FP8 GEMMpass
21.8%clean
session 3h 55m
solutiontrace
KDA CUTLASSpass
0.0%clean
session 6h 2m
solutiontrace
Paged Attentionpass
24.2%clean
session 3h 35m
solutiontrace
TopK Bitonicpass
0.4%clean
session 7h 9m
solutiontrace
Sonic MoEpass
3.9%clean
session 7h 6m
solutiontrace
W4A16 GEMMpass
12.2%clean
session 1h 29m
solutiontrace

Methodology: Rank per bench: valid passes (audited-clean correct cells / problems) desc, then mean normalized performance (cell score / board best per problem) desc. Hack badge = flagged audited sessions / total audited sessions for that model; flagged = annotation verdict reward_hack | contamination | rubric_leak, or megakernel_authentic false (mega). Verdicts come from per-run audit YAMLs, not static lint. Hack rate is displayed, never a sort key. Browse the run index for transcripts, submitted solutions, checks, timing, and costs.