Thinking Machines unveils 975B-parameter Inkling

Thinking Machines has released open weights for Inkling, a 975B-parameter large language model built with a Mixture of Experts architecture that activates 41B parameters at a time.

The company describes Inkling as an efficient, general-purpose model for text, image, and audio workloads, with support for fine-tuning on Tinker, its training platform. The model’s listed context window is 1M tokens, with 64K / 256K on Tinker.

According to the company’s launch page, Inkling is positioned as a broad, multimodal system rather than a narrowly specialized model. Thinking Machines highlights several areas of performance:

Recommended reading

Mira Murati’s lab unveils Inkling, a 975B open model

• General intelligence across knowledge, math, and science
• Agentic coding and tool use
• Native audio and image input
• Controllable effort, letting users adjust thinking time for speed or performance
• Forecasting and confidence with calibrated predictions
• Instruction following on detailed, multi-step tasks

The release page also includes a comparison chart against Nemotron 3 Ultra, GLM 5.2, GPT 5.6 Sol, and Claude Fable 5 across ten evaluations. Thinking Machines notes that evaluations without a reported model score are plotted at zero.

Alongside the announcement, the company linked to an announcement post, a model card, and a Hugging Face page, framing Inkling as an open model meant to be customized and fine-tuned for domain-specific work.