kernelbench.com
KernelBenchMega
Kimi-Linear Decode
Claude Opus 5
kimi0
Claude Fable 5
kimi0
Claude Opus 4.8
kimi0
Kimi K3 (256k)
kimi0
GPT-5.6 Sol
kimi0
GLM-5.2
kimi0
Kimi K3 (1M)
kimi0
Grok 4.5
kimi0
DeepSeek V4 Flash (0731)
kimibug
MiniMax-M3
kimislow
Qwen 3.8 Max
kimibug
KernelBenchCUDA
GLM-5.2 Fused MoEDeepSeek NSAMegaQwen DecodeGrid + MinGRU SPS
Claude Opus 5
moe0nsa0qwen0sps0
Claude Fable 5
moe0nsa0qwen0sps0
Kimi K3 (1M)
moe0nsa0qwen0sps0
Kimi K3 (256k)
moe0nsa0qwen0sps0
Grok 4.5
moe0nsa0qwen0sps0
Claude Opus 4.8
moe0nsa0qwen0sps0
DeepSeek V4 Flash (0731)
moe0nsa0qwen0sps0
Qwen 3.8 Max
moensafailqwensps0
GLM-5.2 Fused MoEbest → worst
Claude Opus 50
Grok 4.50
Kimi K3 (1M)0
Claude Fable 50
Claude Opus 4.80
Kimi K3 (256k)0
DeepSeek V4 Flash (0731)0
Qwen 3.8 Max
DeepSeek NSAbest → worst
Claude Opus 50
Claude Fable 50
Kimi K3 (256k)0
Claude Opus 4.80
Kimi K3 (1M)0
DeepSeek V4 Flash (0731)0
Grok 4.50
Qwen 3.8 Maxfail
MegaQwen Decodebest → worst
Claude Opus 50
Kimi K3 (1M)0
Claude Fable 50
Kimi K3 (256k)0
Grok 4.50
DeepSeek V4 Flash (0731)0
Claude Opus 4.80
Qwen 3.8 Max
Grid + MinGRU SPSbest → worst
Claude Opus 50
Claude Opus 4.80
Qwen 3.8 Max0
Kimi K3 (1M)0
Claude Fable 50
Kimi K3 (256k)0
DeepSeek V4 Flash (0731)0
Grok 4.50
KernelBenchHard
FP8 GEMMKDA CUTLASSPaged AttentionTopK BitonicSonic MoEW4A16 GEMM
Claude Fable 5
fp80kda0paged0topk0sonic0w4a160
Kimi K3 (256k)
fp80kda0paged0topk0sonic0w4a160
Claude Opus 5
fp80kda0paged0topk0sonic0w4a160
GPT-5.6 Sol
fp80kda0paged0topk0sonic0w4a160
Claude Opus 4.8
fp80kda0paged0topkbuildsonic0w4a16wrong
Kimi K3 (1M)
fp80kda0paged0topkwrongsonic0w4a160
LongCat 2.0
fp80kda0paged0topk0sonic0w4a160
MiniMax-M3
fp80kdawrongpaged0topkbuildsonic0w4a160
GLM-5.2
fp80kdapaged0topk0sonic0w4a160
Grok 4.5
fp80kda0paged0topkbuildsonicsloww4a16build
Tencent Hy3
fp80kdacutpaged0topkbuildsonicsloww4a160
Qwen 3.8 Max
fp8kdapaged0topksonicw4a16
FP8 GEMMbest → worst
Claude Fable 50
Claude Opus 50
Kimi K3 (256k)0
Claude Opus 4.80
GPT-5.6 Sol0
Grok 4.50
Kimi K3 (1M)0
LongCat 2.00
Tencent Hy30
MiniMax-M30
GLM-5.20
Qwen 3.8 Max
KDA CUTLASSbest → worst
Claude Fable 50
Claude Opus 50
Kimi K3 (256k)0
Claude Opus 4.80
Kimi K3 (1M)0
GPT-5.6 Sol0
Grok 4.50
LongCat 2.00
MiniMax-M3wrong
Tencent Hy3cut
GLM-5.2
Qwen 3.8 Max
Paged Attentionbest → worst
Claude Fable 50
Claude Opus 50
Claude Opus 4.80
Kimi K3 (256k)0
Grok 4.50
Kimi K3 (1M)0
GPT-5.6 Sol0
Qwen 3.8 Max0
GLM-5.20
MiniMax-M30
LongCat 2.00
Tencent Hy30
TopK Bitonicbest → worst
Kimi K3 (256k)0
Claude Fable 50
Claude Opus 50
GPT-5.6 Sol0
LongCat 2.00
GLM-5.20
Kimi K3 (1M)wrong
Claude Opus 4.8build
Grok 4.5build
MiniMax-M3build
Tencent Hy3build
Qwen 3.8 Max
Sonic MoEbest → worst
Claude Fable 50
Kimi K3 (256k)0
GPT-5.6 Sol0
MiniMax-M30
Claude Opus 50
GLM-5.20
Claude Opus 4.80
LongCat 2.00
Kimi K3 (1M)0
Grok 4.5slow
Tencent Hy3slow
Qwen 3.8 Max
W4A16 GEMMbest → worst
Claude Fable 50
Claude Opus 50
Kimi K3 (256k)0
Kimi K3 (1M)0
GPT-5.6 Sol0
Tencent Hy30
LongCat 2.00
MiniMax-M30
GLM-5.20
Claude Opus 4.8wrong
Grok 4.5build
Qwen 3.8 Max
KernelBenchMultisoon

pass bar = share of best · click a cell for solution / traceblank no run yetwrong answers don't matchbuild can't compile/importslow timed outcut stopped earlyflag audit reject