Thinking Machines

Thinking Machines Lab was founded in February 2025 by Mira Murati (OpenAI's former CTO) with co-founders including John Schulman (co-creator of RLHF and PPO) and roughly two dozen OpenAI alumni. Its $2B seed round at a $12B valuation (July 2025; NVIDIA, Accel, ServiceNow, Cisco, AMD, Jane Street) was the largest in history. A follow-on raise at $50–60B collapsed by January 2026 amid a wave of departures back to OpenAI — co-founders Barret Zoph and Luke Metz among them — after which PyTorch creator Soumith Chintala took over as CTO. The ~150-person lab retains multibillion-dollar compute commitments from NVIDIA and Google.

The lab's thesis is open weights plus customization, against one-size-fits-all frontier APIs. Its first product, Tinker (October 2025), is a managed fine-tuning API that exposes low-level training loops over open-weight models. The flagship Inkling (July 15, 2026) is a from-scratch 975B-parameter (41B active) natively multimodal MoE pretrained on 45T tokens of text, images, audio, and video, released under Apache 2.0 — the leading U.S. open-weights model on the AA Intelligence Index at release, with a controllable thinking-effort dial that reaches peer scores at roughly a third of the tokens.

Its Connectionism research blog punches above its weight: posts on defeating nondeterminism in LLM inference, LoRA fine-tuning dynamics, on-policy distillation, and modular manifolds have been unusually influential for a lab this young, and the accompanying batch-invariant-ops code drew 1K+ GitHub stars.