StorySoftware
Thinking Machines releases Inkling-Small (276B parameters, 12B active) on Tinker and Hugging Face
Thinking Machines released Inkling-Small, a mixture-of-experts model with 276B parameters that uses 12B per token, with full open weights on Hugging Face and fine-tuning on Tinker. It reads text, images and audio with up to a 1M-token context, and on several tests it beats the four-times-larger Inkling at a fraction of the cost.
- SWE-bench Verified 80.2% vs Inkling's 77.6%; Humanity's Last Exam (text) 31.6% vs 29.7%
- Output costs $1.20 per million tokens on Tinker, against $4.05 for Inkling
- Encoder-free design: audio and image patches feed straight into the same model, no separate vision or speech encoder
- Trained on NVIDIA GB300 NVL72 systems