Daily AI CatchupMira Murati's Thinking Machines Releases Inkling, a 975B Open Multimodal Model
Ex-OpenAI CTO Mira Murati's Thinking Machines Lab released Inkling, its first model trained from scratch with full public weights — a 975B-parameter MoE with 41B active params, trained on 45 trillion tokens across text, images, audio, and video with a 1M-token context window. Using an encoder-free architecture, it more than doubles Qwen3-Omni on native audio understanding, framed as a platform play rather than a benchmark chase.
Why it matters
💻 Developer · The encoder-free native audio architecture is worth studying directly if you're building anything audio-heavy — it's a meaningfully different approach than bolting an audio encoder onto a text model.
📦 Product · A credible, fully open frontier-adjacent model from a well-funded new lab gives the U.S. a rare homegrown open-source option in a lane Chinese labs have dominated — worth adding to your model-evaluation shortlist.
🎨 Design · No direct design impact — this is a foundation model release aimed at developers and fine-tuners.
📈 Business · Positioned as a customization platform (via Thinking Machines' Tinker service) rather than a benchmark-chasing release — worth watching if your organization needs a fine-tunable open base model rather than an off-the-shelf API.
🤔 Just Curious · Mira Murati, who used to be OpenAI's chief technology officer, just released her new AI lab's first model — and unlike most big AI companies, she's making the whole thing open and downloadable for anyone.