RUMBA - Russian: leaderboard
RUMBA (ai-forever Russian User Memory Benchmark), Russian split: questions about facts, updates and time spread across long multi-session user dialogues, with the whole history given to the model (full-context setting). Question-weighted LLM-judge accuracy, 0-100, higher is better; RAG and memory-system entries are excluded.
Metric: LLM-judge accuracy (%). Source: huggingface.co. Saturation forecast: Around February 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 83.6 | #83 |
| 2 | Claude Sonnet 4.6 | 78.61 | #94 |
| 3 | Grok 4.1 Fast | 76.22 | #227 |
| 4 | Gemini 3.1 Flash Lite (Preview) | 70.06 | #201 |
| 5 | GPT-4.1 Mini | 60.08 | #363 |
| 6 | minimax-01 | 56.71 | #616 |
| 7 | Llama 4 Maverick | 48.35 | #486 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=rumba-russian · How It Works · Data refreshed daily, snapshot 2026-11-05.