Inkling-Small Preview
Thinking Machines Lab · 2026-07-15
Inkling-Small Preview is the earlier 276B-total, 12B-active checkpoint evaluated in the July 15, 2026 Inkling launch post. Thinking Machines Lab later continued agentic coding reinforcement learning for two weeks before releasing the final Inkling-Small weights, so Atlas keeps the preview checkpoint separate from the July 30 artifact.
Benchmark scores
| Benchmark | Score |
|---|---|
| AIME 2026 | 95.1 |
| Audio MultiChallenge (Text Output) | 49.6 |
| Capability | 152.2 |
| CharXiv Reasoning | 76.7 |
| CharXiv Reasoning (with tools) | 83.4 |
| Global MMLU Lite | 86.8 |
| GPQA Diamond | 88.3 |
| Humanity's Last Exam | 46.6 |
| Humanity's Last Exam (Text-Only) | 29.6 |
| IFBench (Artificial Analysis) | 83.4 |
| MMMU Pro | 73.1 |
| SimpleQA Verified | 20.9 |
| SWE-Bench Pro (Public Dataset) | 53.2 |
| SWE-bench Verified | 77.4 |
| Terminal-Bench 2.1 | 52.7 |
| 𝜏³-Banking | 13.6 |