Thinking Machines unveils Inkling: Open-weights AI model

Two years ago, Meera Murati, formerly CTO of OpenAI, left her position without a backup plan, driven by a desire to explore new ventures. This led to the creation of Thinking Machines, a company that secured a $2 billion investment from A16Z. Until recently, the company was primarily known for its $12 billion valuation before releasing any products.

Introducing Inkling

Thinking Machines has now released Inkling, an open-weights model trained from scratch. Inkling possesses capabilities to see, hear, reason, and can be fine-tuned for real-world applications.

The Genesis of Thinking Machines

In the past year, Meera Murati assembled a formidable team, including OpenAI co-founder John Schulman and VP of Research Barrett Zoph, along with numerous senior researchers. This team first launched Tinker, an API designed for fine-tuning open-weight models. Tinker allows users to customize models without managing their own infrastructure or losing control over the training process.

Inkling's Architecture and Capabilities

Inkling is a mixture-of-experts model featuring 970 billion total parameters. However, it doesn't utilize all parameters simultaneously. Instead, each token is routed to a small subset of specialized experts, meaning only 41 billion parameters are activated per token. This design allows Inkling to achieve the intelligence of a massive model with significantly reduced computational demands.

The model was pre-trained on 45 trillion tokens of text, images, and audio, and it supports a 1 million token context window. Unlike many other open models, Inkling's weights are Apache licensed and are available on Hugging Face.

Performance and Strategic Positioning

On standard benchmarks, Inkling's raw intelligence is surpassed by models like Fable 5 and GPT 5.6 Soul, placing it in the mid-range among open Chinese models. Its release was closely followed by Moonshot's announcement of Kimmy K3, a 2.8 trillion-parameter model that competes directly with Fable 5.

Despite not being the most powerful model in terms of raw intelligence, Inkling's design emphasizes flexibility and efficiency. It features a "thinking effort" dial that allows users to adjust the model's computational intensity. Lowering the dial provides quick, inexpensive answers, while increasing it enables the model to match the performance of Neotron 3 Ultra on Terminal Bench, using only a third of the tokens. This feature is particularly beneficial for agents running millions of times daily.

Innovative Features

One of Inkling's flagship demonstrations involved connecting it to Tinker and instructing it to remove its own ability to use the letter "E." The model then wrote its own training script, generated data, executed the job, and loaded new weights, effectively "lobotomizing" itself.

Inkling also incorporates "epistemics," a training methodology that rewards the model for admitting what it doesn't know rather than making confident but incorrect guesses. This makes Inkling one of the leading models for forecasting future events, outperforming GPT 5.5 and Opus 4.8, and it will explicitly state when it is uncertain.

Furthermore, Inkling processes raw audio and pixels directly, bypassing the need for separate encoder models that translate images and audio into tokens.

Unique Training Side Effects

During its extensive reinforcement learning, specifically after 30 million training rounds, Inkling's internal monologue began to shorten words to save tokens, transitioning from proper English to a more concise, "caveman-like" speech pattern.

The Business Strategy

Thinking Machines acknowledges that Inkling is not intended to dethrone leading AI labs. The core strategy is to provide a capable, free model and then charge for fine-tuning services via Tinker. This allows users to customize Inkling into specialized models tailored to specific problems.