Inkling 2 Mini: benchmark results
Thinking Machines Lab's Inkling 2 Mini (Nov 20, 2026), an Apache 2.0 multimodal looped-transformer MoE (26.4B total, 3.4B unique active / 5.4B effective active per token across 3 looped passes; mixed NVFP4/MXFP8). Provider: Thinking Machines. Released 2026-11-20. Access: Open.
Aggregate Rating 1745 ± 14, rank #187 of 10516 rated models, from 123 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AA-AnalystAgent - Pass@K | 91.25 | Pass@K (%) | 100 |
| LLM Stats (CritPT) | 31.4 | Score (%) | 100 |
| AA-AnalystAgent - Pass@1 | 71 | Pass@1 (%) | 98.4 |
| LLM Stats (LongBench v2) | 66.9 | Score (%) | 94.7 |
| LLM Stats (ScreenSpot Pro) | 76.5 | Score (%) | 87.5 |
| LLM Stats (MCP Atlas) | 82.7 | Score (%) | 82.2 |
| Vals AI CyberBench | 74.58 | Accuracy (%) | 81.5 |
| AA-Briefcase - Presentation Elo | 1440.13 | Elo | 76.8 |
| Vals AI MedScribe | 86.37 | Accuracy (%) | 76.8 |
| AA GDP.pdf - Legal - Criterion Pass Rate | 81.74 | Criterion Pass Rate (%) | 74.6 |
| EnterpriseOps-Gym-AA - Email | 63.46 | Task Success (%) | 74.2 |
| EnterpriseOps-Gym-AA - ITSM | 36.46 | Task Success (%) | 71 |
Interactive version: theaggregate.ai/model?slug=inkling-2-mini · How It Works · Data refreshed daily, snapshot 2026-11-21.