All models
Moonshot AI · 1T MoE · 32B active

Kimi K3

Moonshot's newest mixture-of-experts. Agentic coding and long-context reasoning; runs locally on a 512 GB unified-memory machine.

Verified
Verifiers re-ran it
Pending
No verdict yet
Rejected
Tested, but not proven faster

Kernel Code Efficiency Ranking

Kernels for Kimi K3, ranked by verified speedup. Buy a license for 0.1 SUI: 70% goes to the tuner, 20% to the kernel it improved on, 10% to Opti-om. Open a kernel for how it improved the model, the harness conditions and its contract.

#KernelSpeedupVerified byTuner
1Flash decode · tile 64LLM Inference · 3ab5915609… · sample1.92×1/1 passedincl. platform harness0xab4e…b2db
2Paged attention · simdgroup matmulLLM Inference · eab0b7cd6f… · sample1.79×3/3 passedverified0xe320…3859
3GQA decode · split-KLLM Inference · bab4959534… · sample1.39×4/4 passedverified0x405c…61c6
4Fused RoPE + KV writeLLM Inference · 10d62b2bff… · sample1.38×4/4 passedverified0x9093…3021
5RMSNorm · float4 + warp shuffleLLM Inference · 2869af1423… · sample1.21×1/1 passedincl. platform harness0x5566…afa0
#KernelSpeedupVerified byTuner
1Long-context KV compactionCoding · cc203c1903… · sample2.45×3/3 passedverified0xec00…6a9f
2Prefix-cache attentionCoding · 090e542f2a… · sample2.34×4/4 passedincl. platform harness0x24f2…0151
3Speculative decode · 4-token draftCoding · dd3983c84f… · sample2.04×4/4 passedverified0x4aeb…a60c
4RMSNorm · float4 + warp shuffleCoding · b30b75f4a2… · sample1.52×4/4 passedverified0xf607…1bca
Nothing verified for Kimi K3 on Image Gen yet. The first kernel verifiers confirm takes #1.
#KernelSpeedupVerified byTuner
1Motion vectors · warp-coherentVideo · b10f60fcab… · sample2.45×3/3 passedverified0x7305…3287
2SAD 16×16 · shared-memory tilesVideo · a0508c64a0… · sample2.20×2/2 passedverified0x5e92…f434
—Frame delta · int8Video · d3ee0c2eda… · sample1.62×1/2 passedincl. platform harness0x5d02…0536
Nothing verified for Kimi K3 on 3D Rendering yet. The first kernel verifiers confirm takes #1.
#KernelSpeedupVerified byTuner
1FP64 GEMM · tensor-core emulationScientific / HPC · a8bff7324d… · sample2.09×3/3 passedincl. platform harness0x48d6…1f69
2SpMV CSR · merge-path load balanceScientific / HPC · 1535633d7a… · sample1.96×2/2 passedverified0x0297…654d
3Stencil 7-pt · shared tilesScientific / HPC · 4157396ca0… · sample1.92×2/2 passedverified0x24ab…2101

↑ ↓ to move between workloads · Esc for all · Sample rows are generated; rows marked verified on this platform come from real submissions and can be licensed.