Inkling: Our open-weights model

Jul 15, 2026

Our model, called Inkling, is a Mixture-of-Experts transformer with 975B total parameters, 41B active. It supports a context window of up to 1M tokens. It was pretrained on 45 trillion tokens of text, images, audio and video. It is the first in a family of models of different sizes: alongside it we are sharing a preview of Inkling-Small, a lighter-weight model with 12B active parameters, trained with a similar recipe, that achieves strong performance with even lower cost and latency.

Inkling reasons natively over text, images, and audio, and balances cost with performance through efficient and controllable thinking effort. We trained it to be a broad, balanced foundation model: strong across many domains, flexible enough to adapt. Inkling is not the strongest overall model available today, open or closed. Instead, a combination of qualities makes it a good open-weights base for customization: multimodal capabilities, efficient thinking, and availability for fine-tuning.

We want to make customization accessible for more use cases, so Inkling is available for fine-tuning. Picking the right base model to fine-tune is a qualitative judgment that combines measurable benchmarks with the unique feel of a model that comes from playing with it.

To show what customization means in practice, we asked Inkling to fine-tune itself. Using Tinker, it produced and ran its own fine-tuning job:

inkling@tinker: self-finetuning
who are you?
I’m Inkling, an AI assistant running thinkingmachines/Inkling. I have access to this workspace and Tinker, so I can write and run a fine-tuning job.
How can I help with the model?  
fine tune yourself with Tinker to be a lipogram model where you never use the letter e in your responses.
I need to handle the full loop: create a training dataset and objective, fine-tune the current model on Tinker, evaluate it against the base model, and switch to the improved weights.

Capabilities

Generalist models trained across a variety of tasks allow for customization that meets real-world needs. Inkling can adapt to various workflows and applications, demonstrating strong performance on benchmarks for trustworthiness and safety.

Generalist model

Inkling is designed to be broad, trained across agentic, reasoning, coding, instruction-following, factuality, vision, and audio tasks. A spider chart compares Inkling with Nemotron 3 Ultra, GLM 5.2, GPT 5.6 Sol, and Claude Fable 5 on performance evaluations.

Inkling is a broad, balanced generalist model. Benchmark scores reflect competitive performance across text, agentic, multimodal, and audio evaluations.

Agentic coding and tool use

Inkling is trained for flexibility in solving a wide variety of tasks with agentic tool use. Demonstrations showcase its capabilities in building a functional web app, and creating personalized interfaces.

Controllable thinking effort

Inkling supports controllable thinking effort, allowing developers to balance performance with token efficiency. The ability to adjust effort via system messages ensures suitability for various tasks across applications.

The making of Inkling

Inkling's architecture is built on a Mixture-of-Experts transformer framework, utilizing a unique setup tailored for efficiency and long-context performance. Trained on 45 trillion tokens of diverse content, the model was developed to exhibit robustness and versatility across audio, visual, and textual inputs.

Inkling-Small

Alongside Inkling, we previewed Inkling-Small, matching or exceeding performance on many benchmarks but optimized for lower latency and cost-efficiency.

Customizing Inkling

Inkling is readily customizable, enabling rapid learning through fine-tuning to adapt to specialized organizational knowledge and specific requirements.

Inkling availability

Inkling is available on Tinker, with discounts for a limited time and options for context lengths of 64K and 256K tokens. We continue to partner with various platforms to facilitate the deployment of Inkling for diverse use cases.