Thinking Machines Lab releases open-weight model Inkling
The 975-billion-parameter mixture-of-experts model was pitched not as the strongest available but as a base for enterprise fine-tuning through the lab's Tinker platform.
- Models & capabilities
- Open weights & ecosystem
- Minor
Thinking Machines Lab, the company Mira Murati founded after leaving OpenAI, released its first in-house model, an open-weight mixture-of-experts system called Inkling. It has 975 billion total parameters with roughly 41 billion active per task, trained from scratch on 45 trillion tokens of text, image, audio and video, though current outputs are limited to text, code, styled artefacts and structured data.
The lab was explicit that Inkling was “not the strongest overall model available today, open or closed,” according to TechCrunch’s report. Its case instead was that broadly capable open weights, adaptable through the company’s existing Tinker fine-tuning platform, would beat general-purpose closed models for specific organisational uses — citing a project with Bridgewater Associates that reportedly reached 84.7% accuracy on financial-reasoning tests at a fraction of the inference cost of proprietary alternatives. For post-training, Thinking Machines used other open-weight models, including Moonshot AI’s Kimi K2.5, to bootstrap early data before its own reinforcement learning, saying a future model would use “fully self-contained post-training instead.”
The release positioned Thinking Machines, founded in 2025, as a full-stack competitor to Meta and the Chinese open-weight labs rather than purely a tooling company, betting that customisability would matter more than leaderboard position.