MiniCPM-SALA: benchmark results
OpenBMB's 9B model hybridizing sparse and linear attention for million-token context on consumer GPUs (February 2026). Provider: OpenBMB. Released 2026-02-11. Access: Open.
Unified ELO 1497 ± 1, rank #702 of 1392 rated models, from 11 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| HELMET (128K) | 45.4 | Average Score | 83.1 |
| LLM Stats (AIME 2024) | 83.75 | Score (%) | 63.5 |
| OpenCompass Research - AIME 2025 | 81 | Score (%) | 42.6 |
| LLM Stats Score | 14.44 | LLM Stats Score (conservative rating) | 36.8 |
| OpenCompass Research - HLE | 9.3 | Score (%) | 34 |
| OpenCompass Research - LiveCodeBench v6 | 54.7 | Score (%) | 29.8 |
| OpenCompass | 58.7 | Average Score | 21.3 |
| LLM Stats (CMMLU) | 81.55 | Score (%) | 20 |
| OpenCompass Research - MMLU-Pro | 72 | Score (%) | 8.5 |
| OpenCompass Research - GPQA Diamond | 58.2 | Score (%) | 4.3 |
| OpenCompass Research - IFEval | 76.9 | Score (%) | 2.1 |
Interactive version: theaggregate.ai/model?slug=minicpm-sala · How It Works · Data refreshed daily, snapshot 2026-09-05.