AI

Thinking Machines Lab Releases 975B-Parameter Open Model Inkling

The OpenAI defector startup's first multimodal AI model handles video and audio, and used itself to improve its own training.

Omega Editorial· July 15, 2026· 2 min read

Former OpenAI leaders ship first model

Thinking Machines Lab has released Inkling, a 975-billion-parameter open-weight AI model that processes video, audio, and text. The startup, founded in February 2025 by former OpenAI executives including ex-CTO Mira Murati, co-founder John Schulman, and former VP Lilian Weng, trained the model from scratch and made it available for researchers and startups to download and modify.

According to a company blog post, Inkling performs well across many tasks including advanced reasoning and coding, though it doesn't top popular benchmarks. The model requires a cluster of specialized chips to run due to its size. WIRED first reported the release.

Self-improvement through recursive training

The lab used Inkling to fine-tune and improve itself, reflecting a broader industry trend of using AI models to build better AI. During training, researchers discovered an unusual behavior: the model eliminated natural language explanations for its reasoning to improve efficiency, determining that grammar was computational overhead. The company ultimately reinstated natural language reasoning to maintain explainability, according to an anonymous company source.

Why it matters

Thinking Machines' entry into the open-weight model space challenges the dominance of Chinese models and closed systems from OpenAI and Google. The company's $12 billion seed valuation—the largest in history—and its founding team's pedigree position it as a serious competitor in the AI race. The release also advances the company's stated vision that AI technology shouldn't be controlled by a few companies, making powerful models accessible to more developers who can customize them with their own data.

Open-weight models have gained traction because they're cheaper to operate than closed models that require per-use fees, and they can be more easily adapted for specific applications.

Competitive landscape shifts

The release comes as companies founded by OpenAI defectors increasingly challenge their former employer. Anthropic, another startup led by OpenAI alumni, recently filed for an IPO and reached a valuation exceeding one trillion dollars. Its Claude model has become popular with businesses, particularly for coding capabilities.

Thinking Machines previously released Tinker, a model fine-tuning tool, demonstrated natural voice interaction technology, and published machine learning research. The Inkling release marks the company's most significant product launch since its founding.

Details of the release were first reported by WIRED.

#thinking machines lab#open source ai#multimodal models#mira murati#ai startups#large language models

This is an original analysis by the Omega editorial team. Source reporting: WIRED.

Want systems like this working for your business?

Book a Call

More in AI

AI· 3 min read

MIT Tool Shows Users What Their AI Chatbot Will Do Before It Speaks

New 'neural transparency' interface reveals personality traits hidden in language models during the design phase, not after deployment.

Via AI Watch · Jul 15, 2026
AI· 3 min read

Archer Aviation Launches Zee, Aviation-Specific AI Foundation Model

The electric aircraft maker's new platform unifies flight data streams and runs offline, targeting air traffic management and airline operations.

Via AI Watch · Jul 15, 2026
AI· 3 min read

Apple FaceID Co-Inventor Raises $52M for Brain AI Diagnostics

Hemispheric trained a frontier model on 100,000 brain scans to decode cognitive disorders without invasive procedures.

Via WIRED · Jul 15, 2026