Thinking Machines released Inkling-Small, a freely downloadable model that matches or beats its larger Inkling model on reasoning and coding tests while using about a quarter of the size and under a third of the price per unit of output. It handles text, images and audio, and the amount of reasoning it spends per request can be dialled up or down. The larger Inkling still leads on factual knowledge accuracy.
What changed
Thinking Machines' only released model was Inkling, roughly four times larger and about three times the price per unit of output.
What it unlocks
Running or fine-tuning a model that handles text, images and audio on far less hardware, with downloadable weights and adjustable amounts of reasoning per request.
- 276B total parameters, 12B active, versus 975B total and 41B active for Inkling
- output price $1.20 per 1M tokens versus $4.05 for Inkling
- context window up to 1M tokens
- 31.6% on Humanity's Last Exam, above Inkling's 29.7%
- 80.2% on SWEBench Verified
What you need to act on it
- hardware to run the downloaded weights, or a Tinker account for fine-tuning and chat
- thinkingmachines.ai2026-07-30