Inkling AI model launches from Mira Murati’s Thinking Machines

The Inkling AI model is the first major release from Mira Murati’s Thinking Machines Lab, and it gives developers a new open-source option at a time when many leading open models have come from Chinese AI labs. Thinking Machines released Inkling on July 15 with full model weights on Hugging Face under an Apache 2.0 license, according to Decrypt.

For investors watching the AI race, the release is another sign that model access, pricing and developer adoption are becoming as important as headline performance. Open-source AI models can be downloaded, inspected and built into products more freely than closed systems, which can lower the barrier for startups and independent developers.

What is the Inkling AI model?

Inkling is a 975-billion-parameter mixture-of-experts model trained from scratch by Thinking Machines Lab. A parameter is a learned setting inside an AI model, while a token is a unit of information the model reads or writes, such as part of a word.

The model uses 41 billion active parameters during inference, the stage when a model generates an answer. Decrypt reported that Inkling can process text, images and audio, supports a 1-million-token context window, and was pretrained on 45 trillion tokens.

A mixture-of-experts model routes tasks through selected parts of a larger model rather than using every parameter for every answer. That can make very large models more practical to run, although Decrypt noted that Inkling is still not something ordinary users should expect to operate locally on their own hardware.

Where Inkling appears strongest

Thinking Machines’ benchmark data points to agentic tool use as Inkling’s clearest strength. Agentic AI refers to systems that do multi-step tasks through tools, rather than only responding in chat.

On MCP Atlas, a benchmark that measures how reliably an AI agent completes real-world tasks through the Model Context Protocol standard, Inkling scored 74.1%, according to Thinking Machines data cited by Decrypt. Decrypt said that result was nearly 30 percentage points higher than Nvidia’s Nemotron 3 Ultra.

Inkling also scored 77.6% on SWE-Bench Verified, a benchmark for autonomous GitHub bug fixing, according to the same report. That was above Nemotron’s 70.7% score.

Those numbers suggest Inkling may be more useful for workflows where a model has to operate tools, edit code or complete structured tasks. Benchmark scores are controlled tests, so they do not guarantee how the model will perform in every real-world setup.

How much does Inkling cost?

Inkling is available on OpenRouter at $1 per million input tokens and $4.05 per million output tokens, according to OpenRouter pricing cited by Decrypt. Input tokens are the material a user sends to the model, while output tokens are the answer the model produces.

Decrypt reported that Hermes and OpenClaw setups using OpenRouter can switch to Inkling without additional configuration. That availability could help developers test the model without hosting the full system themselves.

In hands-on testing, Decrypt said Inkling showed mixed results. The model said through Thinking Machines’ own interface that it was fully private, according to the review. On coding tasks, Decrypt reported that a complex 1,955-word prompt for a zombie-shooting game produced a blank screen, while a shorter 99-word version led the model to produce a working game.

Decrypt’s overall read was that Inkling is a serious Western open-source entrant, especially for agent-style workflows, while Chinese models still appeared stronger on raw coding performance for the cost.

This story draws on original reporting from Decrypt.