Inkling: Thinking Machines Lab ships its first open-weights model - a 975B MoE, Apache-2.0, multimodal

The first open-weights model from Thinking Machines Lab (Mira Murati) is here: Inkling, MoE 975B params (41B active), multimodal, Apache-2.0 license. The most discreet box in the frontier scene enters the open market through the front door.

The fact

On July 16, Simon Willison shared on his blog the release of Inkling, presented as the first open-weights model from Thinking Machines Lab (the structure of Mira Murati, former CTO of OpenAI). Announced specs: Mixture-of-Experts transformer, 975 B parameters in total (41 B active), Apache-2.0 license, multimodal (text, image, audio, video), trained on 45 T tokens.

Our analysis

This is a major development in an already crowded open landscape (Kimi K3, DeepSeek, Qwen3, GLM-4.6), but with two singularities: the Apache-2.0 license (permissive, unrestricted commercial exploitation, to be confirmed on all modules) and the origin (Thinking Machines, hitherto silent since its record fundraising). The choice of open by a structure raised to frontier valuation is not a technical choice - it's a distribution choice. Facing Meta which subsidizes Llama and Mistral which hesitates, Thinking Machines frames its entry.

To watch

Position on LMSys and LiveCodeBench within 10 days; real inference footprint (41 B active = accessible on a single high-end GPU?); adoption by orchestrators (Fireworks, Together, DeepInfra); first third-party multimodal fine-tunes; implicit business model (managed services, enterprise, or pure gift to the commons).