Thinking Machines Lab

Thinking Machines Lab is an AI research company founded in February 2025 by Mira Murati, formerly chief technology officer of OpenAI, and staffed heavily with researchers drawn from OpenAI, Google DeepMind, and Meta. It operates out of San Francisco with roughly 150 employees, and counts Nvidia, AMD, Cisco, and Jane Street among its backers alongside lead investor Andreessen Horowitz.

The lab's thesis is that the useful frontier is customization rather than raw capability. Where most frontier labs sell access to a single closed model tuned to be good at everything, Thinking Machines argues that organizations get more out of a strong general base they can adapt to their own data and their own evaluation criteria. Its two products follow directly from that position: Tinker, an API for fine-tuning language models that deliberately exposes control over training algorithms and data, and Inkling, the lab's first in-house model.

Inkling is a multimodal mixture-of-experts (MoE) system with 975 billion total parameters that activates roughly 41 billion for any given token, trained on 45 trillion tokens spanning text, images, audio, and video, with a context window reaching one million tokens. It ships under an Apache 2.0 license with weights hosted on Hugging Face. The lab is unusually direct about where the model lands, stating plainly that Inkling is not the strongest overall model available today, open or closed, and positioning it as a base for customization rather than a leaderboard entry. A smaller preview variant, Inkling-Small, activates 12 billion of 276 billion total parameters at lower latency and cost.

That candor is strategically consistent rather than modest. If the product is the customization layer, the base model does not need to win benchmarks outright — it needs to be strong, open, and adaptable enough that a fine-tuned version beats a closed general-purpose model on the narrow task a customer actually cares about. The open question is whether enough organizations have the appetite and the applied talent to do that work, or whether most default to renting a closed frontier model instead.