Thinking Machines open-sources Inkling-Small, a 12B-active MoE
Open weights near the full Inkling at a quarter the size: native audio and images, a 1M-token context, 80% on SWEBench, plus a Gemini robotics API.
Copy markdown
Frontier-class scores at a quarter the weights
Inkling-Small is a 276B-parameter MoE with only 12B active, yet it edges the full Inkling on Humanity's Last Exam (31.6% vs 29.7%) and posts 80.2% on SWEBench Verified. Open weights are on Hugging Face, so you can self-host a near-frontier coder instead of renting one.
Native audio and images, not bolted on
One model reasons over text, images, and audio directly — audio as dMel spectrograms, images as 40x40 patches — with up to a 1M-token context. That's an open base for voice agents, screen-understanding tools, and long-document pipelines without stitching three models together.
12B active means it fits a real budget
Because only 12B parameters fire per token, latency and inference cost track a mid-size model while quality tracks the big one. Thinking Machines pitches it as a drop-in for coding and agent workloads at roughly a quarter of full Inkling's serving cost.
Fine-tune it on your own data through Tinker
Beyond raw weights, you can fine-tune Inkling-Small on Thinking Machines' Tinker platform and try it in the Tinker Playground before committing. Open weights plus a managed fine-tune path is rare for a model this capable.
Elsewhere: Gemini Robotics ER 2 opens up as an API
Google's Gemini Robotics ER 2 is now in preview via the Gemini API and AI Studio, exposing spatial reasoning, video progress-tracking, and tool-orchestration you can call like any other Gemini model — a door into physical-AI agents with no robotics stack of your own.