Inkling is a flagship, open-weight artificial intelligence model developed by Thinking Machines Lab, a startup founded by former OpenAI CTO Mira Murati. It is a massive model designed as a "balanced generalist" across reasoning, coding, and multimodal tasks. [1, 2, 3]
Inkling: Our Open-Weights Model - Thinking Machines Lab
Architecture & Parameters: A Mixture-of-Experts (MoE) transformer featuring 975 billion total parameters and 41 billion active parameters, supporting up to a 1M token context window.
Multimodal Capabilities: Pretrained on 45 trillion tokens of text, images, audio, and video using an encoder-free architecture for native vision and speech processing.
Controllable Thinking Effort: Allows developers to adjust the model's reasoning effort (from 0.2 to 0.99) to balance performance against cost and latency.
Agentic Tool Use & Coding: Performs strongly on agentic benchmarks (such as SWEBench and Terminal Bench) and supports browser use, code generation, and multi-step tasks.
Calibrated Epistemics: Trained using RL against proper scoring rules to improve factual accuracy, calibration (confidence estimation), and resistance to censorship without over-refusing benign prompts.
Customization & Availability: Released with open weights on Hugging Face and integrated into the Tinker platform for fine-tuning, alongside a preview of a lighter Inkling-Small model (12B active parameters).
