✳ the wire · intel
Inkling-Small is out — 276B/12B active, 80.2% SWE-bench Verified, Apache 2.0, and our catalogs never saw it
Thinking Machines released Inkling-Small on July 30, closing the loop on the preview that shipped alongside Inkling two weeks ago. It comes off our watchlist and onto the board as live.
The specs: 276B total / 12B active MoE, natively multimodal across text, images and audio, 512K-token context, trained on NVIDIA GB300 NVL72. Benchmarks from the model card — SWE-bench Verified 80.2%, GPQA Diamond 89.5%, IFBench 82.2%, MMMU Pro (Vision) 74.0%, Humanity's Last Exam 31.6%, SciCode 48.7%. An NVFP4 checkpoint ships alongside the full weights on Hugging Face.
Every wire item is source-verified before it posts and graded: confirmed · strong · reported · rumor. Subscribers see the primary source on each one, the reasoning behind its grade, and the full analysis — plus the launch alerts themselves, which fire the minute a model answers rather than whenever the feed catches up.