Inkling: benchmark results
Thinking Machines Lab's first open-weights model, a 975B-A41B multimodal MoE with controllable thinking effort. Provider: Thinking Machines. Released 2026-07-15. Access: Open.
Aggregate Rating 1741 ± 7, rank #126 of 1656 rated models, from 538 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Nejumi 4 - jaster (0-shot) - MMLU-ProX (Japanese) | 93 | Exact match (%) | 100 |
| PACT - Rule-Scope Discernment | 0.91 | Score (0-1) | 100 |
| Nejumi 4 - jaster (2-shot) - JSQuAD | 82.28 | Character F1 (x100) | 99.3 |
| Enkrypt AI - Jailbreak Risk | 0.72 | Risk Score | 98.8 |
| Nejumi 4 - GLP - Expression | 99.33 | Score (%) | 98.5 |
| AA Omniscience - Software Engineering (SWE) - C | 75 | Accuracy (%) | 98 |
| Nejumi 4 - M-IFEval (Japanese) | 92.04 | Score (%) | 97.9 |
| Nejumi 4 - ALT - Controllability | 93.54 | Score (%) | 97.8 |
| Nejumi 4 - GLP - Applied Language | 90.62 | Score (%) | 97.8 |
| AA Omniscience - Software Engineering (SWE) - HTML | 68 | Accuracy (%) | 97.4 |
| Nejumi 4 - GLP - Logical Reasoning | 96.5 | Score (%) | 97.4 |
| AA Omniscience - Software Engineering (SWE) - TypeScript | 60 | Accuracy (%) | 97.3 |
Interactive version: theaggregate.ai/model?slug=inkling · How It Works · Data refreshed daily, snapshot 2026-10-20.