Latest30 Jul 2026Open-weights LLM
Inkling-Small
A smaller sibling of Inkling with 276B total and 12B active parameters that takes text, image and audio input natively. Thinking Machines says it performs comparably to Inkling at a quarter of the size, and reports 80.2% on SWE-bench Verified. It was previewed with Inkling and released two weeks later under Apache 2.0.
- 276B total / 12B active MoE
- Up to 1M-token context
- Text, image and audio input; text output
- Lab-stated SWE-bench Verified 80.2%, GPQA Diamond 89.5%
AccessOpen weights on Hugging Face (thinkingmachines/Inkling-Small, Apache 2.0); fine-tuning and playground on Tinker.
15 Jul 2026Open-weights LLM
Inkling
Thinking Machines Lab's first model: a 975B-parameter mixture-of-experts with 41B active parameters, trained on 45 trillion tokens of text, images, audio and video, and released as open weights. The lab presents it as a balanced generalist and a base for fine-tuning on Tinker, and says it is not the strongest model available today.
- 975B total / 41B active MoE
- Up to 1M-token context (64K/256K on Tinker)
- Text, image, audio and video input
- Lab-stated SWE-bench Verified 77.6%, AIME 2026 97.1%
- NVFP4 checkpoint for NVIDIA Blackwell
AccessOpen weights on Hugging Face (thinkingmachines/inkling); fine-tuning on Tinker; hosted APIs via Together AI, Fireworks, Modal, Databricks and Baseten.