Thinking Machines Lab

Artificial Intelligence · Open-Source AI · Enterprise Software

Launch

Thinking Machines releases Inkling, its first open-weights AI model

The 975B-parameter model bets enterprises want AI they can customize rather than rent from frontier labs.

• Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, released Inkling on July 15, 2026 — its first proprietary AI model, with full weights open for anyone to download and modify.
• The company's flagship product is Tinker, launched October 2025, a platform for fine-tuning AI models; Inkling is designed as the customizable base model that powers it.
• Inkling is a Mixture-of-Experts transformer with 975B total parameters (41B active), a 1M-token context window, pretrained on 45 trillion tokens spanning text, images, audio, and video.
• A smaller preview version, Inkling-Small, uses 12B active parameters and in some benchmarks like IFBench (83.4% vs 79.8%) actually outperforms its larger sibling.
• On one benchmark, Inkling matches Nvidia's Nemotron 3 Ultra coding performance using roughly a third as many tokens, thanks to a controllable 'thinking effort' setting.
• Thinking Machines explicitly says Inkling is 'not the strongest overall model available today, open or closed,' positioning it instead as the best available base for customization.
• Analysis platform Artificial Analysis found Inkling has a 63% hallucination rate and calls it the most powerful U.S. open-weights model, though it still trails top Chinese open models like Kimi K2.6 and DeepSeek v4.
• Inkling arrives as Western open-weight AI has ceded ground to Chinese labs following Meta's retreat from open models after Llama 4, making a credible U.S. alternative for enterprise customization strategically significant.