LongCat Flash Lite: benchmark results
Meituan non-thinking MoE with 68.5B total parameters, including over 30B embedding parameters. Dynamic activation ranges from 2.9B to 4.5B, described as approximately 3B. Supports 256K context through YaRN. Provider: Meituan. Released 2026-01-29. Access: Open.
Aggregate Rating 1531 ± 37, rank #681 of 1656 rated models, from 54 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Tau2-Bench Telecom | 79.53 | Success Rate (%) | 70.7 |
| AA TAU-2 Bench | 79.53 | Accuracy (%) | 69.7 |
| ZeroEval MATH-500 | 96.8 | MATH-500 Score | 68.8 |
| AA Omniscience - Software Engineering (SWE) - Dart | 18 | Accuracy (%) | 51.5 |
| AA Omniscience - Software Engineering (SWE) - Kotlin | 16 | Accuracy (%) | 47.8 |
| Artificial Analysis Intelligence Index | 11.47 | Intelligence Index | 46.8 |
| AA IFBench | 43.06 | Accuracy (%) | 46.3 |
| AA Omniscience - Software Engineering (SWE) - Julia | 8 | Accuracy (%) | 46.1 |
| AA Terminal-Bench Hard | 10.61 | Accuracy (%) | 45.1 |
| AA Omniscience - Software Engineering (SWE) - PHP | 18 | Accuracy (%) | 43 |
| LLM Stats Score | 18.33 | LLM Stats Score (conservative rating) | 40.3 |
| LLM Stats (CMMLU) | 82.48 | Score (%) | 40 |
Interactive version: theaggregate.ai/model?slug=longcat-flash-lite · How It Works · Data refreshed daily, snapshot 2026-10-20.