kernelbench.com
KernelBenchMega
Kimi-Linear Decode
Claude Opus 5.5
kimi0
Claude Sonnet 5.5
kimi0
Gemini 4 Argon (High)
kimi0
GPT-6 Astra Pro
kimi0
Claude Fable 5
kimi0
Claude Fable 5.1
kimi0
GLM-5.3
kimi0
Kimi K3 (256k)
kimi0
DeepSeek V4.1 Flash
kimi0
GLM-5.3 Flash
kimi0
Grok 4.7
kimi0
GPT-6.1 Sol
kimi0
Gemini 3.8 Flash (High)
kimi0
GPT-5.6 Sol
kimi0
GPT-6 Luna
kimi0
Muse Spark 1.3
kimi0
GPT-6 Sol
kimi0
KernelBenchCUDA
GLM-5.2 Fused MoEDeepSeek NSAMegaQwen DecodeGrid + MinGRU SPS
Claude Sonnet 5.5
moe0nsa0qwen0sps0
Claude Opus 5.5
moe0nsa0qwen0sps0
Gemini 4 Argon (High)
moe0nsa0qwen0sps0
Claude Fable 5.1
moe0nsaqwen0sps0
Claude Fable 5
moe0nsa0qwen0sps0
DeepSeek V4.1 Flash
moe0nsa0qwen0sps
Claude Opus 5
moensaqwen0sps0
Grok 4.7
moe0nsa0qwen0sps0
GPT-6 Astra Pro
moe0nsaqwen0sps0
GPT-6.1 Sol
moe0nsa0qwen0sps0
GPT-6 Sol
moe0nsa0qwen0sps0
Claude Haiku 5.5
moe0nsa0qwen0sps0
Grok 4.6
moe0nsaqwen0sps
GPT-6 Luna
moe0nsa0qwen0sps0
Muse Spark 1.3
moe0nsa0qwen0sps
Kimi K3 (256k)
moe0nsaqwen0sps0
Qwen 3.8 Max
moe0nsaqwensps0
GLM-5.3
moe0nsaqwensps
DeepSeek V4 Flash (0731)
moe0nsaqwen0sps0
Gemini 3.8 Flash (High)
moensa0qwen0sps0
GLM-5.3 Flash
moensa0qwen0sps
GLM-5.2 Fused MoEbest → worst
Claude Sonnet 5.50
Claude Opus 5.50
Claude Fable 5.10
Qwen 3.8 Max0
GLM-5.30
Gemini 4 Argon (High)0
Grok 4.70
Claude Haiku 5.50
DeepSeek V4.1 Flash0
GPT-6 Astra Pro0
GPT-6.1 Sol0
Grok 4.60
GPT-6 Sol0
Muse Spark 1.30
Claude Fable 50
GPT-6 Luna0
Kimi K3 (256k)0
DeepSeek V4 Flash (0731)0
DeepSeek NSAbest → worst
Claude Sonnet 5.50
Claude Opus 5.50
Claude Fable 50
Gemini 4 Argon (High)0
DeepSeek V4.1 Flash0
GPT-6.1 Sol0
Claude Haiku 5.50
Grok 4.70
Gemini 3.8 Flash (High)0
GLM-5.3 Flash0
GPT-6 Luna0
GPT-6 Sol0
Muse Spark 1.30
MegaQwen Decodebest → worst
Claude Sonnet 5.50
Claude Opus 5.50
Claude Opus 50
Claude Fable 5.10
Claude Fable 50
Grok 4.60
Gemini 4 Argon (High)0
DeepSeek V4.1 Flash0
GPT-6 Astra Pro0
Kimi K3 (256k)0
Claude Haiku 5.50
Grok 4.70
Gemini 3.8 Flash (High)0
GPT-6 Sol0
GLM-5.3 Flash0
GPT-6.1 Sol0
Muse Spark 1.30
GPT-6 Luna0
DeepSeek V4 Flash (0731)0
Grid + MinGRU SPSbest → worst
Claude Opus 50
Claude Sonnet 5.50
Claude Opus 5.50
Claude Fable 5.10
GPT-6 Astra Pro0
GPT-6 Sol0
Grok 4.70
GPT-6 Luna0
Gemini 4 Argon (High)0
Gemini 3.8 Flash (High)0
Claude Haiku 5.50
GPT-6.1 Sol0
Qwen 3.8 Max0
Claude Fable 50
Kimi K3 (256k)0
DeepSeek V4 Flash (0731)0
KernelBenchHard
FP8 GEMMKDA CUTLASSPaged AttentionTopK BitonicSonic MoEW4A16 GEMM
DeepSeek V4.1 Flash
fp80kda0paged0topk0sonic0w4a160
Grok 4.6
fp80kda0paged0topk0sonicw4a160
Claude Fable 5
fp80kdapaged0topk0sonic0w4a160
Kimi K3 (256k)
fp80kda0paged0topk0sonic0w4a16
Qwen 3.8 Max
fp80kdapaged0topk0sonic0w4a16
GLM-5.3
fp8kdapaged0topk0sonicw4a160
DeepSeek V4 Flash (0731)
fp80kdapaged0topk0sonicw4a160
Claude Opus 5
fp8kdapagedtopk0sonicw4a160
GPT-5.6 Sol
fp80kdapagedtopk0sonicw4a160
GLM-5.3 Flash
fp8kdapagedtopk0sonic0w4a16
FP8 GEMMbest → worst
DeepSeek V4 Flash (0731)0
Claude Fable 50
GPT-5.6 Sol0
Grok 4.60
DeepSeek V4.1 Flash0
Qwen 3.8 Max0
Kimi K3 (256k)0
KDA CUTLASSbest → worst
Grok 4.60
DeepSeek V4.1 Flash0
Kimi K3 (256k)0
Paged Attentionbest → worst
Qwen 3.8 Max0
GLM-5.30
Grok 4.60
DeepSeek V4 Flash (0731)0
Kimi K3 (256k)0
Claude Fable 50
DeepSeek V4.1 Flash0
TopK Bitonicbest → worst
Claude Opus 50
GLM-5.30
GLM-5.3 Flash0
Kimi K3 (256k)0
Claude Fable 50
DeepSeek V4.1 Flash0
GPT-5.6 Sol0
Qwen 3.8 Max0
Grok 4.60
DeepSeek V4 Flash (0731)0
Sonic MoEbest → worst
DeepSeek V4.1 Flash0
Qwen 3.8 Max0
Claude Fable 50
GLM-5.3 Flash0
Kimi K3 (256k)0
W4A16 GEMMbest → worst
Claude Opus 50
GLM-5.30
Claude Fable 50
DeepSeek V4.1 Flash0
Grok 4.60
GPT-5.6 Sol0
DeepSeek V4 Flash (0731)0
KernelBenchMultisoon

pass bar = share of best · click a cell for solution / traceblank no audited pass (failed, unaudited or not run; the model page lists every outcome)