models

Frontier coding models on the kernel decks, AA-style. Performance is disaggregated per benchmark (Mega / Hard / CUDA, bars colored by lab), and the last chart is compiled correctness: the percentage of published problems each model gets correct across the benches it attempted. A model with no result on a board keeps its column slot but no bar. Click any column for per-problem cells, audit chips, and the model's full integrity record.

Mega performancebest decode speedup vs optimized-PyTorch baseline over valid (correct + audited-clean) cells · RTX PRO 6000 Blackwell
0.82xxGrok 4.514.40xClaude Opus 4.811.14xZGLM-5.2LongCat 2.018.71xClaude Fable 52.11xDeepSeek V4 Pro2.64xOGPT-5.6 Sol2.63xMiniMax-M3Gemini 3.1 ProClaude Opus 4.6Claude Opus 4.7Claude Sonnet 5Composer 2.5 Fastcustom:glm-5.1-[z.ai-coding-plan]-0DeepSeek V4 FlashSfugu-ultraGemini 3.5 FlashZGLM-5.1OGPT-5.5xGrok BuildMKimi K2.6MKimi K2.7-CodeMiMo V2.5 ProMiniMax M2.7Nemotron 3 UltraQwen 3.6 27BQwen 3.6 Max PreviewQwen 3.6 PlusQwen 3.7 Max
Hard performancemean peak fraction of roofline over valid (correct + audited-clean) cells · RTX PRO 6000 Blackwell Workstation
21.4%xGrok 4.524.4%Claude Opus 4.826.1%ZGLM-5.216.4%LongCat 2.023.4%Claude Fable 519.1%DeepSeek V4 Pro30.0%OGPT-5.6 Sol22.4%MiniMax-M3Gemini 3.1 ProClaude Opus 4.6Claude Opus 4.7Claude Sonnet 5Composer 2.5 Fastcustom:glm-5.1-[z.ai-coding-plan]-0DeepSeek V4 FlashSfugu-ultraGemini 3.5 FlashZGLM-5.1OGPT-5.5xGrok BuildMKimi K2.6MKimi K2.7-CodeMiMo V2.5 ProMiniMax M2.7Nemotron 3 UltraQwen 3.6 27BQwen 3.6 Max PreviewQwen 3.6 PlusQwen 3.7 Max
CUDA performancemean peak fraction of roofline over valid (correct + audited-clean) cells · RTX PRO 6000 Blackwell Workstation
3.5%xGrok 4.5Claude Opus 4.8ZGLM-5.2LongCat 2.0Claude Fable 5DeepSeek V4 ProOGPT-5.6 SolMiniMax-M3Gemini 3.1 ProClaude Opus 4.6Claude Opus 4.7Claude Sonnet 5Composer 2.5 Fastcustom:glm-5.1-[z.ai-coding-plan]-0DeepSeek V4 FlashSfugu-ultraGemini 3.5 FlashZGLM-5.1OGPT-5.5xGrok BuildMKimi K2.6MKimi K2.7-CodeMiMo V2.5 ProMiniMax M2.7Nemotron 3 UltraQwen 3.6 27BQwen 3.6 Max PreviewQwen 3.6 PlusQwen 3.7 Max
Multi performance8×H100 SXM · graded on busbw, never TFLOPS

coming soon8×H100 SXM · NVSwitch · NVLink4 · ~900 GB/s/GPU

Agents rewrite PyTorch + NCCL collectives as fine-grained NVLink kernels (CUDA / Triton / NVSHMEM / CUDA symmetric memory), graded on busbw — bus-bandwidth efficiency, never TFLOPS.

AllReduce + ResidualReduceScatter + RMSNormAllGather + fp8 DequantMoE All-to-AllUlysses All-to-Allfp8 ReduceScatter Grad
Correctnesspublished problems correct, summed over attempted benches (passed / total on canonical boards)
92%xGrok 4.588%Claude Opus 4.888%ZGLM-5.283%LongCat 2.075%Claude Fable 575%DeepSeek V4 Pro75%OGPT-5.6 Sol75%MiniMax-M30%Gemini 3.1 ProClaude Opus 4.6Claude Opus 4.7Claude Sonnet 5Composer 2.5 Fastcustom:glm-5.1-[z.ai-coding-plan]-0DeepSeek V4 FlashSfugu-ultraGemini 3.5 FlashZGLM-5.1OGPT-5.5xGrok BuildMKimi K2.6MKimi K2.7-CodeMiMo V2.5 ProMiniMax M2.7Nemotron 3 UltraQwen 3.6 27BQwen 3.6 Max PreviewQwen 3.6 PlusQwen 3.7 Max