Inkling 2 Small Launch - Sparse Attention Kernel Speedup: leaderboard
Preliminary DeepSeek-V3.2 sparse-attention kernel optimization, forward and backward. Best of four independent sessions on one B300 GPU. Curves show correctness-passing benchmark records; endpoint markers show final speedups verified with independent timing on hidden tests. Speedup is original kernel time divided by optimized kernel time. Claude Opus 5.5: 18.80× at 8.875 hours; GPT-6 Astra: 10.26× at 3.692 hours; Inkling 2 Small: 6.02× at 5.988 hours; DeepSeek V4.1 Flash: 5.66× at 2.764 hours; GLM 5.3: 5.03× at 6.921 hours; Mistral Large 4: 3.43× at 5.498 hours. These are kernel-only speedups, not end-to-end training gains; numerical correctness is tolerance-checked, not bit-exact. The task prompt has no time limit. Endpoint times differ; do not label the outcomes as a fixed equal-time comparison or end-to-end training speedup. Native final verified endpoints are used, not an inferred value from SVG coordinates. Model effort is not specified for this experiment.
Metric: Verified kernel speedup (times original). Source: thinkingmachines.ai. Saturation forecast: Rough model projection: around 2026. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5.5 (Benchmark Sweep Run: source=Inkling2SmallKernel; best of 4 sessions; 1xB300; final independently verified endpoint) | 18.8 | |
| 2 | GPT-6 Astra (Benchmark Sweep Run: source=Inkling2SmallKernel; best of 4 sessions; 1xB300; final independently verified endpoint) | 10.26 | |
| 3 | Inkling 2 Small (Benchmark Sweep Run: source=Inkling2SmallKernel; best of 4 sessions; 1xB300; final independently verified endpoint) | 6.02 | |
| 4 | DeepSeek V4.1 Flash (Benchmark Sweep Run: source=Inkling2SmallKernel; best of 4 sessions; 1xB300; final independently verified endpoint) | 5.66 | |
| 5 | GLM 5.3 (Benchmark Sweep Run: source=Inkling2SmallKernel; best of 4 sessions; 1xB300; final independently verified endpoint) | 5.03 | |
| 6 | Mistral Large 4 (Benchmark Sweep Run: source=Inkling2SmallKernel; best of 4 sessions; 1xB300; final independently verified endpoint) | 3.43 |
Interactive version: theaggregate.ai/benchmark?slug=inkling-2-small-launch-sparse-attention-kernel-speedup · How It Works · Data refreshed daily, snapshot 2026-11-06.