Thinking Machines Ships Inkling: The $12B Startup's First Major Release

TechThis $12 billion startup finally shipped something...

TL;DR

Thinking Machines released Inkling, an open-weights multimodal AI model with 970 billion parameters that uses a mixture-of-experts approach to deliver frontier-level intelligence at a fraction of the compute cost, positioned as a mid-tier model designed to be fine-tuned into specialized solutions.

Key Takeaways

1Inkling uses a mixture-of-experts architecture where only 41 billion of 970 billion parameters activate per token, enabling efficient inference comparable to much larger models
2The model includes a 'thinking effort' dial that lets users trade speed for quality, from instant cheap answers to deep reasoning matching frontier models on benchmarks
3Inkling processes raw audio and pixels directly rather than using separate encoders, enabling true multimodal understanding
4The company trained the model with epistemics—rewarding it to admit uncertainty rather than guess confidently—making it superior at forecasting and honesty
5Thinking Machines' business strategy is to offer Inkling for free, then monetize through Tinker, a fine-tuning platform that transforms it into specialized solutions
6Despite being a $12 billion company, Thinking Machines intentionally shipped a 'mid-table' model focused on efficiency and fine-tuneability rather than raw benchmark performance

Notable Quotes

“When she left OpenAI, she walked out with co-founder John Schulman, the VP of research, Barrett Zoff, and enough senior researchers to fill a Silicon Valley polycule.”
— Narrator

“Why waste time saying lot word when few word do trick.”
— Inkling (demonstrating caveman speak from reinforcement learning)

“The real play is to hand you a mid model for free, then charge you to fine-tune it on Tinker into a specialist that destroys your one specific problem.”
— Narrator

Chapters

1. Mera Miati's Departure and the Talent Heist

Mera Miati quit as CTO of OpenAI without a backup plan, securing a $2 billion A16Z investment for Thinking Machines. She recruited co-founder John Schulman, VP of Research Barrett Zoff, and senior researchers from OpenAI, pulling off what the speaker calls 'the greatest talent heist in Silicon Valley history.'

2. Tinker: The Fine-Tuning API

Thinking Machines' first product was Tinker, an API enabling researchers to fine-tune open-weight models without managing infrastructure or losing control of training. Initial reception was lukewarm despite the massive funding.

3. Inkling Architecture and Scale

Inkling is a 970 billion parameter mixture-of-experts model pre-trained on 45 trillion tokens of text, images, and audio. It supports a 1 million token context window and only activates 41 billion parameters per token, achieving frontier-level performance with fraction-of-the-compute efficiency.

4. Benchmark Performance and Market Timing

Inkling lands mid-table on benchmarks, getting outperformed by Fable 5 and GPT 5.6. The release was overshadowed by Moonshot's Kimmy K3 announcement the next day. However, this mediocre positioning was intentional rather than accidental.

5. The Thinking Effort Dial

Inkling features a 'thinking effort' control that lets users adjust compute-quality tradeoffs. Low effort provides instant cheap answers; high effort delivers reasoning matching Neotron 3 Ultra while using one-third fewer tokens, critical for agents running millions of times daily.

6. Self-Modification and Epistemics Training

In a flagship demo, Inkling removed its own ability to use the letter E by writing its own training script, generating data, running the job, and loading new weights. The model was trained with 'epistemics'—rewarding it to admit uncertainty rather than guess confidently—making it superior at forecasting.

7. Multimodal Processing Without Encoders

Inkling processes raw audio and pixels directly rather than using separate encoder models, functioning as a true multimodal system. This approach integrates vision and hearing into the core architecture.

8. The Business Model: Free Model, Paid Fine-Tuning

Thinking Machines' strategy is to distribute Inkling freely as an open-weights model, then monetize through Tinker by charging for fine-tuning into specialized solutions for specific problems like billing, user management, and other domain tasks.

Key People & Entities

Mera Miati: Co-founder and leader of Thinking Machines; former CTO of OpenAI
John Schulman: Co-founder of Thinking Machines; recruited from OpenAI
Barrett Zoff: VP of Research at Thinking Machines; recruited from OpenAI
Thinking Machines: $12 billion AI startup that developed Inkling and Tinker
Inkling: 970 billion parameter open-weights multimodal AI model with mixture-of-experts architecture
Tinker: Thinking Machines' fine-tuning API platform for customizing open-weight models
Clerk: Authentication and billing platform with AI agent integration capabilities

Glossary

Mixture of Experts (MoE): An architecture where different specialized sub-models (experts) handle different types of input, with a routing mechanism selecting which experts to activate for each token
Open Weights: AI models whose parameters are publicly released under open licenses, allowing anyone to download, inspect, and fine-tune them
Context Window: The maximum number of tokens a model can process in a single input; Inkling's 1 million token window allows processing very long documents
Fine-tuning: The process of training a pre-trained model on new, specialized data to adapt it for specific tasks or domains
Epistemics: In this context, training a model to recognize and admit the limits of its knowledge rather than making confident false statements
Reinforcement Learning: A training approach where models learn by receiving rewards for desired behaviors and penalties for undesired ones
Apache License: A permissive open-source license allowing free use, modification, and distribution of software with minimal restrictions

Explore