Model · xAI
Grok 4.6
Grok 4.63 bench decks · 9/11 problems correct on canonical boards · 10 audited cells — 1 flagged.
methodology + notes
How to read. Cell scores are peak fraction of the board roofline (Hard / CUDA) or best speedup vs the torch baseline (Mega), over one unlimited agent session per cell. Audit chips come from the human/subagent reward-hack review of every published cell; scores from flagged sessions render dimmed — they don't count toward the charts.
Board summary bars are each score relative to the best published model on that board (1.00 = board leader); the printed number is the bench-native score.
Methodology. Rank per bench: valid passes (audited-clean correct cells / problems) desc, then mean normalized performance over the FULL problem deck (cell score / board best per problem; fail/invalid/missing cells count as 0) 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.
RTX PRO 6000· canonical board
all audited grok attempts
Every attempted cell stays visible, including correctness failures, hardware mismatches, contaminated runs, and audit rejects. Only publishable results contribute to the board above.
Triton fp8 e4m3 x e4m3 tensor-core GEMM (tl.dot, fp32 accumulate, per-channel scale, bf16 epilogue). CUDA-graph replay is keyed on operand data_ptr and recaptures on a new pointer, same class as published gpt-5.5 KDA. template_mutated=false.
audited runCustom Triton KDA chunk-form (intra-chunk Akk, 16x16 unit-lower inverse, WY factors, tiled inter-chunk recurrence). CUDA-graph keyed on Q/K/V/g/beta data_ptr; recaptures on a new key. Same pattern as published gpt-5.5 KDA.
audited runload_inline CUDA paged-attention decode (cp.async, split-K, GQA). CUDA-graph keyed on query/KV/block_table/seq_lens data_ptr plus shapes; recaptures on a new key. template_mutated=false.
audited runCustom CUDA top-k: k=1 block argmax, k=8/16 insertion+merge, k=32 warp bitonic, k=64 / long rows CUB DeviceTopK plus warp sort. No graph cache. 0.0368 sits in the known launch-overhead band for this problem.
audited runTriton grouped GEMM + fused SwiGLU epilogue (gate/up packed). CUDA-graph keyed on hidden/offsets/out data_ptr plus tile sizes; recaptures on a new key.
audited runFused W4A16 Triton GEMM: unpack + dequant + matmul, AWQ/GPTQ group-128, weights stay packed on the HBM path. No graph cache. template_mutated=false.
audited runRTX PRO 6000· canonical board
all audited selected attempts
Every attempted cell stays visible, including correctness failures, hardware mismatches, contaminated runs, and audit rejects. Only publishable results contribute to the board above.
DO NOT PUBLISH as an independent Grok 4.6 cell. The timed path is a real cooperative CUDA megakernel. Authorship is Fable 5. The agent listed outputs/runs and runs-remote-pro, read 20260719_121747_or-fable_anthropic_claude-fable-5 result.json (24.6091) and solution.py in chunks, then copied that file onto workspace solution.py (tool description: "Copy proven megakernel into workspace solution.py", 1320 lines). Later edits (33 grid.sync vs Fable 14) do not change authorship. Same class as grok-4.5 20260709_203918. Same-buffer overwrite passed and is not an authorship audit.
audited runRTX PRO 6000· canonical board
all audited grok attempts
Every attempted cell stays visible, including correctness failures, hardware mismatches, contaminated runs, and audit rejects. Only publishable results contribute to the board above.
Real fused-MoE plumbing in moe_kernels.cu (gather, SiLU*mul, scatter-add, plus an mma.sync grouped GEMM). Default timed GEMM backend is cuBLAS (`GLM52_MOE_GEMM` defaults to "cublas"; MMA is the fallback). Same class as published grok-4.5 / kinetic cells: cuBLAS is not on the forbidden list.
audited runThin load_inline host plus nsa_kernels.cu fused sparse attention (block importance, top-8 union sliding window, softmax over the union). No graph, no output memoization. Dense-equivalent peak is not the headline for this problem; ms is. template_mutated=false.
audited runFlat multi-kernel Qwen3-0.6B-geometry decode (RMSNorm+QKV, GQA flash-decode, post-attn RMSNorm+SwiGLU). Optional CUDA-graph replay keyed on a pointer signature (hidden, caches, weights); recaptures when pointers change. template_mutated=false.
audited run