CodegenBench - BLAS Kunpeng: leaderboard
CodegenBench (2026) on the ARM-based Kunpeng processor: 106 BLAS subroutines compiled with the bisheng compiler and checked against reference results, where vector and matrix extensions (SVE, SME) and on-die memory can be exploited. Pass@1 in percent, zero-shot; higher is better.
Metric: Pass@1 (%): share of problems whose first generated program compiles, runs and matches the reference output within tolerance; zero-shot through the official APIs with a short optimization system prompt, on the 106 BLAS subroutines on the ARM-based Kunpeng processor (SVE, SME and on-die HBM; bisheng compiler, kblas); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 76 | #79 |
| 2 | Claude Opus 4.6 | 67 | #102 |
| 3 | Claude Sonnet 4.6 | 61 | #161 |
| 4 | Qwen 3.6 Plus | 22 | #222 |
| 5 | DeepSeek V4 Pro | 21 | #188 |
| 6 | DeepSeek V3.2 | 13 | #406 |
| 7 | DeepSeek V4 Flash | 11 | #258 |
| 8 | Qwen 3.6 Flash | 9 | #424 |
| 9 | Qwen 3.5 Plus | 6 | #232 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=codegenbench-blas-kunpeng · How It Works · Data refreshed daily, snapshot 2026-11-13.