80% on SWE-Bench Verified From 12B Active Parameters, Apache 2.0

Inkling-Small is Thinking Machines Lab’s first open-weight release: 276B total parameters with just 12B active, 42 decoder layers routing each token to 6 of 256 experts plus 2 shared, hybrid local-global attention, and text, image and audio inputs producing text out. Apache 2.0. The claimed numbers are startling for the active-parameter count — 80.2% SWE-Bench Verified, 89.5% GPQA Diamond, 95.5% AIME 2026. They are also, so far, self-reported and drawing surprisingly little community traction relative to what they assert. Treat as provisional until independent evaluations land, then reassess quickly if they hold.