Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, releases first general model Inkling
Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has released its first general model Inkling, with a total of 975 billion parameters and 41 billion activated per instance.
The model supports text, images, and audio, with a maximum context of 1 million tokens, using MoE architecture, primarily referencing DeepSeek-V3, and trained with open data such as Kimi K2.5.
The complete weights are available on Hugging Face under the Apache 2.0 license, currently focusing on customization and efficiency, with token consumption about one-third that of Nvidia Nemotron 3 Ultra at the same performance level in Terminal Bench 2.1.
Source: Public information
ABAB AI Insight
After leaving OpenAI, Mira Murati quickly established the lab, and the release of Inkling continues her pursuit of efficient large models, reflecting the new layout of former OpenAI core members in the open-source arena.
The lab trained the MoE model from scratch, leveraging open data and architectural reference resources, motivated to create a cost-effective customizable alternative to avoid direct competition with closed giants.
Similar to the DeepSeek series' open-source efficient path, Thinking Machines is currently in the validation phase of its first model, rapidly obtaining community feedback and iteration through the Apache license.
Essentially a technological alternative, the MoE architecture and efficient activation reduce inference costs, shifting the general model focus from parameter competition to performance optimization per token, reconstructing the accessibility threshold of open-source AI.
ABAB News · Cognitive Laws
Activation parameters determine costs, activation efficiency determines victory or defeat.
Open-source weights are the key to community accelerators.
Frontline leavers often reshape the next generation paradigm.