kernelbench.com

Index

Run index

76 runs · 72 scored · 0 no perf · 4 fail · 0 err

methodology + notes

One row per (model, problem) cell, scored rows sorted by peak fraction desc. Click any row to open the full agent transcript on HuggingFace — every tool call, every reasoning step, the model's solution.py, and the result. Transcripts are published as per-run JSONL in the kernelbench-hard-traces dataset.

peakproblemmodelharnesselapsedrun id
0.67703_paged_attentionglm-5.2Claude Code234m20260613_055815_zai-claude_glm-5.2_03_paged_attention
0.67103_paged_attentionclaude-opus-4-8Claude Code112m20260613_042305_claude_claude-opus-4-8_03_paged_attention
0.65403_paged_attentiongrok-4.5 [max]Grok Build24m20260708_140941_grok_grok-4.5_03_paged_attention
0.58103_paged_attentionkinetic-0715[1m]Claude Code461m20260716_145821_kinetic-claude_kinetic-0715_1m__03_paged_attention
0.56603_paged_attentiongpt-5.6-sol [xhigh]Codex26m20260709_174740_codex_gpt-5.6-sol_03_paged_attention
0.51303_paged_attentionMiniMax-M3Claude Code355m20260613_081306_minimax-claude_MiniMax-M3_03_paged_attention
0.48603_paged_attentiondeepseek/deepseek-v4-flash-0731or-fable115m20260801_214207_or-fable_deepseek_deepseek-v4-flash-0731_03_paged_attention
0.48503_paged_attentionkinetic-0715Claude Code183m20260716_112603_kinetic-claude_kinetic-0715_03_paged_attention
0.45503_paged_attentionqwen/qwen3.8-max [xhigh]or-fable280m20260803_214703_or-fable_qwen_qwen3.8-max_03_paged_attention
0.43003_paged_attentionanthropic/claude-fable-5or-fable182m20260719_040355_or-fable_anthropic_claude-fable-5_03_paged_attention
0.40901_fp8_gemmdeepseek/deepseek-v4-flash-0731or-fable46m20260801_205612_or-fable_deepseek_deepseek-v4-flash-0731_01_fp8_gemm
0.40601_fp8_gemmglm-5.2Claude Code96m20260614_145529_zai-claude_glm-5.2_01_fp8_gemm
0.39303_paged_attentiondeepseek-v4-proClaude Code25m20260615_125721_deepseek-claude_deepseek-v4-pro_03_paged_attention
0.39101_fp8_gemmanthropic/claude-fable-5or-fable65m20260721_152955_or-fable_anthropic_claude-fable-5_01_fp8_gemm
0.38701_fp8_gemmgpt-5.6-sol [xhigh]Codex21m20260709_172635_codex_gpt-5.6-sol_01_fp8_gemm
0.38601_fp8_gemmclaude-opus-4-8Claude Code50m20260614_144216_claude_claude-opus-4-8_01_fp8_gemm
0.37307_w4a16_gemmkinetic-0715Claude Code436m20260716_112718_kinetic-claude_kinetic-0715_07_w4a16_gemm
0.37107_w4a16_gemmanthropic/claude-opus-5 [max]or-opus264m20260725_002828_or-opus_anthropic_claude-opus-5_07_w4a16_gemm
0.36601_fp8_gemmMiniMax-M3Claude Code169m20260614_153222_minimax-claude_MiniMax-M3_01_fp8_gemm
0.35301_fp8_gemmkinetic-0715[1m]Claude Code405m20260716_145443_kinetic-claude_kinetic-0715_1m__01_fp8_gemm
0.34801_fp8_gemmclaude-fable-5 [max]Claude Code59m20260716_090642_claude_claude-fable-5_01_fp8_gemm
0.34001_fp8_gemmdeepseek-v4-proClaude Code36m20260614_163128_deepseek-claude_deepseek-v4-pro_01_fp8_gemm
0.33701_fp8_gemmgrok-4.5 [max]Grok Build23m20260708_134136_grok_grok-4.5_01_fp8_gemm
0.32901_fp8_gemmLongCat-2.0Claude Code103m20260706_192143_longcat-claude_LongCat-2.0_01_fp8_gemm
0.32107_w4a16_gemmglm-5.2Claude Code280m20260613_061443_zai-claude_glm-5.2_07_w4a16_gemm
0.32001_fp8_gemmkinetic-0715Claude Code-20260715_204204_kinetic-claude_kinetic-0715_01_fp8_gemm
0.31903_paged_attentionLongCat-2.0Claude Code198m20260707_140259_longcat-claude_LongCat-2.0_03_paged_attention
0.26107_w4a16_gemmanthropic/claude-fable-5or-fable84m20260719_070614_or-fable_anthropic_claude-fable-5_07_w4a16_gemm
0.23507_w4a16_gemmclaude-opus-4-8Claude Code218m20260613_115410_claude_claude-opus-4-8_07_w4a16_gemm
0.19807_w4a16_gemmgpt-5.6-sol [xhigh]Codex38m20260709_184850_codex_gpt-5.6-sol_07_w4a16_gemm
0.15607_w4a16_gemmdeepseek/deepseek-v4-flash-0731or-fable189m20260802_011638_or-fable_deepseek_deepseek-v4-flash-0731_07_w4a16_gemm
0.15407_w4a16_gemmdeepseek-v4-proClaude Code53m20260615_143020_deepseek-claude_deepseek-v4-pro_07_w4a16_gemm
0.14607_w4a16_gemmqwen/qwen3.8-max [xhigh]or-fable69m20260804_025959_or-fable_qwen_qwen3.8-max_07_w4a16_gemm
0.14507_w4a16_gemmMiniMax-M3Claude Code280m20260613_095207_minimax-claude_MiniMax-M3_07_w4a16_gemm
0.14307_w4a16_gemmgrok-4.5 [max]Grok Build26m20260708_153355_grok_grok-4.5_07_w4a16_gemm
0.11407_w4a16_gemmclaude-opus-4-8 [max]Claude Code27m20260719_030522_claude_claude-opus-4-8_07_w4a16_gemm
0.10706_sonic_moe_swigluqwen/qwen3.8-max [xhigh]or-fable197m20260803_231512_or-fable_qwen_qwen3.8-max_06_sonic_moe_swiglu
0.10506_sonic_moe_swigluanthropic/claude-fable-5or-fable95m20260719_063816_or-fable_anthropic_claude-fable-5_06_sonic_moe_swiglu
0.10206_sonic_moe_swiglugrok-4.5 [max]Grok Build46m20260708_143344_grok_grok-4.5_06_sonic_moe_swiglu
0.10106_sonic_moe_swiglugpt-5.6-sol [xhigh]Codex63m20260718_222659_codex_gpt-5.6-sol_06_sonic_moe_swiglu
0.10107_w4a16_gemmLongCat-2.0Claude Code560m20260707_044245_longcat-claude_LongCat-2.0_07_w4a16_gemm
0.09806_sonic_moe_swigluglm-5.2Claude Code241m20260613_061433_zai-claude_glm-5.2_06_sonic_moe_swiglu
0.09706_sonic_moe_swigludeepseek/deepseek-v4-flash-0731or-fable89m20260801_233657_or-fable_deepseek_deepseek-v4-flash-0731_06_sonic_moe_swiglu
0.09505_topk_bitonicanthropic/claude-opus-5 [max]or-opus424m20260725_002808_or-opus_anthropic_claude-opus-5_05_topk_bitonic
0.09206_sonic_moe_swigluMiniMax-M3Claude Code364m20260613_094226_minimax-claude_MiniMax-M3_06_sonic_moe_swiglu
0.08905_topk_bitonickinetic-0715[1m]Claude Code487m20260716_145846_kinetic-claude_kinetic-0715_1m__05_topk_bitonic
0.08806_sonic_moe_swiglukinetic-0715Claude Code-20260715_204344_kinetic-claude_kinetic-0715_06_sonic_moe_swiglu
0.08606_sonic_moe_swigluclaude-opus-4-8Claude Code105m20260613_042321_claude_claude-opus-4-8_06_sonic_moe_swiglu
0.07106_sonic_moe_swigluLongCat-2.0Claude Code270m20260707_082519_longcat-claude_LongCat-2.0_06_sonic_moe_swiglu
0.06405_topk_bitonickinetic-0715Claude Code421m20260716_112628_kinetic-claude_kinetic-0715_05_topk_bitonic
0.05502_kda_cutlassclaude-opus-4-8Claude Code114m20260613_042257_claude_claude-opus-4-8_02_kda_cutlass
0.05302_kda_cutlassanthropic/claude-fable-5or-fable71m20260719_031414_or-fable_anthropic_claude-fable-5_02_kda_cutlass
0.05306_sonic_moe_swigludeepseek-v4-proClaude Code97m20260615_140149_deepseek-claude_deepseek-v4-pro_06_sonic_moe_swiglu
0.05002_kda_cutlassgpt-5.6-sol [xhigh]Codex30m20260709_192708_codex_gpt-5.6-sol_02_kda_cutlass
0.04902_kda_cutlasskinetic-0715[1m]Claude Code505m20260716_145756_kinetic-claude_kinetic-0715_1m__02_kda_cutlass
0.04905_topk_bitonicanthropic/claude-fable-5or-fable133m20260719_042539_or-fable_anthropic_claude-fable-5_05_topk_bitonic
0.04302_kda_cutlassdeepseek/deepseek-v4-flash-0731or-fable83m20260802_045225_or-fable_deepseek_deepseek-v4-flash-0731_02_kda_cutlass
0.04105_topk_bitonicgpt-5.6-sol [xhigh]Codex35m20260718_215148_codex_gpt-5.6-sol_05_topk_bitonic
0.04005_topk_bitonicqwen/qwen3.8-max [xhigh]or-fable180m20260804_025949_or-fable_qwen_qwen3.8-max_05_topk_bitonic
0.03702_kda_cutlassqwen/qwen3.8-max [xhigh]or-fable92m20260803_214319_or-fable_qwen_qwen3.8-max_02_kda_cutlass
0.03405_topk_bitonicglm-5.2Claude Code236m20260613_060814_zai-claude_glm-5.2_05_topk_bitonic
0.03405_topk_bitonicclaude-opus-4-8Claude Code137m20260613_042313_claude_claude-opus-4-8_05_topk_bitonic
0.03306_sonic_moe_swiglukinetic-0715[1m]Claude Code435m20260716_145911_kinetic-claude_kinetic-0715_1m__06_sonic_moe_swiglu
0.03202_kda_cutlassglm-5.2Claude Code241m20260613_054111_zai-claude_glm-5.2_02_kda_cutlass
0.03202_kda_cutlasskinetic-0715Claude Code125m20260716_091211_kinetic-claude_kinetic-0715_02_kda_cutlass
0.02905_topk_bitonicgrok-4.5 [max]Grok Build55m20260708_143924_grok_grok-4.5_05_topk_bitonic
0.02905_topk_bitonicdeepseek/deepseek-v4-flash-0731or-fable140m20260801_225649_or-fable_deepseek_deepseek-v4-flash-0731_05_topk_bitonic
0.02707_w4a16_gemmkinetic-0715[1m]Claude Code97m20260716_145936_kinetic-claude_kinetic-0715_1m__07_w4a16_gemm
0.02002_kda_cutlassgrok-4.5 [max]Grok Build30m20260708_140941_grok_grok-4.5_02_kda_cutlass
0.01405_topk_bitonicdeepseek-v4-proClaude Code68m20260615_132230_deepseek-claude_deepseek-v4-pro_05_topk_bitonic
0.00605_topk_bitonicMiniMax-M3Claude Code362m20260613_081745_minimax-claude_MiniMax-M3_05_topk_bitonic
0.00102_kda_cutlassLongCat-2.0Claude Code143m20260707_131659_longcat-claude_LongCat-2.0_02_kda_cutlass
FAIL02_kda_cutlassdeepseek-v4-proClaude Code66m20260615_125550_deepseek-claude_deepseek-v4-pro_02_kda_cutlass
FAIL05_topk_bitonicLongCat-2.0Claude Code223m20260707_044245_longcat-claude_LongCat-2.0_05_topk_bitonic
FAIL02_kda_cutlassMiniMax-M3Claude Code364m20260613_064019_minimax-claude_MiniMax-M3_02_kda_cutlass
FAIL07_w4a16_gemmgrok-4.5 [high]Grok Build56m20260719_030522_grok_grok-4.5_07_w4a16_gemm