A.I.
More labs
Inkling
Inkling is the first model from Thinking Machines Lab. It was released on 15 July 2026 with full open weights on Hugging Face (Thinking Machines, 15 Jul 2026). The lab pitches it as a base to customize with Tinker, its fine-tuning service. Fine-tuning is extra training on your own data.
- Status
- Current
- Released
- 15 Jul 2026
- Last checked
- 12 Oct 2026newest source date on this page
Key facts
| Field | Inkling |
|---|---|
| Size | 975 billion parameters, 41 billion active |
| Context | Up to 1M tokens; 64K and 256K options on Tinker |
| Inputs | Text, images, and audio |
| Training data | 45 trillion tokens of text, images, audio, and video |
| Training hardware | Nvidia GB300 NVL72 |
| Smaller sibling | Inkling-Small: 276 billion parameters, 12 billion active |
Thinking Machines says plainly that Inkling is "not the strongest overall model available today, open or closed." It is also served by Together AI, Fireworks, Modal, Databricks, and Baseten.
What the lab reports
- 77.6% on SWE-bench Verified, a coding test, with a bash-only harness, is the lab's own score; 87.2% on GPQA Diamond, a science test, is Artificial Analysis's score, as quoted by the lab. Both at effort 0.99.