Thinking Machines Lab Releases Inkling, an Open-Weight 975B Parameter MoE Model

Thinking Machines Lab, the AI startup led by former OpenAI chief technology officer Mira Murati, released its first model on July 15, 2026. Inkling is an open-weight mixture-of-experts model available on Hugging Face and the company's Tinker platform.

Inkling has 975 billion total parameters with 41 billion active per forward pass. It was trained on 45 trillion tokens and supports a 1 million token context window. The model handles text, image, and audio inputs through a sliding window multimodal architecture.

A Model Built for Breadth, Not Specialization

Inkling is designed to perform well across a wide range of tasks rather than excel at a single benchmark. The company describes it as a foundation model for practical use and customization. Key architectural choices include:

• Mixture-of-experts with DeepSeek auxiliary-free load balancing and two shared experts instead of the typical one.
• Relative position bias instead of rotary position embeddings, with small convolutional layers in several places.
• Controllable thinking effort, allowing users to adjust how much computation the model applies to a given query.

The model was pretrained from scratch starting last winter, with post-training beginning in mid-January. The team scaled reinforcement learning runs to over 30 million rollouts and thousands of continuous training hours. John Schulman, a co-founder of Thinking Machines Lab, said a small team built up coding, reasoning, and agentic training from there.

Implications for the Open-Source AI Landscape

Inkling is notable for being independently pretrained, not distilled from larger frontier models. Martin Casado, a general partner at Andreessen Horowitz, called it the only highly capable frontier model not distilled from the big labs. The model was trained on NVIDIA GB300 NVL72 hardware with NVFP4 checkpoints for efficient inference.

The release comes as open-weight models from Chinese labs like DeepSeek and Qwen have gained traction. Thinking Machines Lab positions Inkling as a Western alternative with a focus on resistance to censorship. The model is available for fine-tuning on Tinker and through Databricks. The company plans to continue iterating on Inkling and expanding its ecosystem of tools and integrations.