DeepSeek V4 Pro

DeepSeek11 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

deepseek-claude/deepseek-v4-pro5/60.538 audited

RTX PRO 6000

FP8 GEMMpass
34.0%clean
session 36m
solutiontrace
KDA CUTLASSfail
bug
session 1h 6m
solutiontrace
Paged Attentionpass
39.3%clean
session 25m
solutiontrace
TopK Bitonicpass
1.4%clean
session 1h 8m
solutiontrace
Sonic MoEpass
5.3%clean
session 1h 37m
solutiontrace
W4A16 GEMMpass
15.4%clean
session 53m
solutiontrace

B200

FP8 GEMMpass
5.9%unaudited
session 51m
solutiontrace
KDA CUTLASSpass
0.0%unaudited
session 48m
solutiontrace
Paged Attentionpass
6.3%unaudited
session 47m
solutiontrace
TopK Bitonicfail
unaudited
session 45m
solutiontrace
Sonic MoEpass
4.0%unaudited
session 49m
solutiontrace
W4A16 GEMMfail
unaudited
session 45m
trace

H100 PCIe

FP8 GEMMpass
5.9%unaudited
session 49m
solutiontrace
KDA CUTLASSfail
unaudited
session 57m
solutiontrace
Paged Attentionpass
1.5%unaudited
session 54m
solutiontrace
TopK Bitonicpass
0.7%unaudited
session 46m
solutiontrace
Sonic MoEpass
5.5%unaudited
session 52m
solutiontrace
W4A16 GEMMpass
3.7%unaudited
session 49m
solutiontrace

RTX 3090

FP8 GEMMno run
KDA CUTLASSpass
0.5%unaudited
session 31m
trace
Paged Attentionpass
3.7%unaudited
session 2h 17m
trace
TopK Bitonicpass
0.2%unaudited
session 1h 16m
trace
Sonic MoEfail
unaudited
session 1h 2m
trace
W4A16 GEMMpass
16.7%unaudited
session 20m
trace

mega

deepseek-claude/deepseek-v4-pro1/20.113 audited

RTX PRO 6000

RL Grid PPOno run
Kimi-Linear Decodepass
2.11xclean
439 tok/s  2048 ctx 2.42x8192 ctx 2.16x16384 ctx 1.79x · triton
solutiontrace

B200

RL Grid PPOno run
Kimi-Linear Decodepass
1.56xclean
148 tok/s  2048 ctx 1.56x8192 ctx 1.57x16384 ctx 1.56x · triton
solutiontrace

H100 PCIe

RL Grid PPOno run
Kimi-Linear Decodepass
1.38xclean
178 tok/s  2048 ctx 1.59x8192 ctx 1.40x16384 ctx 1.19x · triton
solutiontrace

Legacy pre-v2 hard board: best 6/8 passed across snapshot labels opencode/deepseek/deepseek-v4-pro.

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.