Inkling is the first model released by Mira Murati’s Thinking Machines Lab, and the source describes it as the best open-source model trained from scratch by a Western lab, a description used in coverage. The model has 975 billion total parameters, with 41 billion active at inference, and its trained weights are reported as available on Hugging Face under an Apache 2.0 license.
Inkling uses a mixture-of-experts architecture. The model is reported to have 975 billion total parameters, with 41 billion parameters active during inference. It accepts text, images, and audio as input modalities and supports a one-million-token context window. These architectural and input-capability attributes are presented as part of the model’s reported specifications.
Inkling was pretrained on 45 trillion tokens. The reported figures separately list the architecture type, parameter counts, context window size, and pretraining token count as discrete attributes. The provided facts do not include further dataset names, preprocessing steps, or additional training regimen details. The description here mirrors only the reported technical architecture and training scale as included in the supplied material. All items above are stated in the provided material without further operational details.
Inkling’s MCP Atlas score is 74.1%. Its SWE-Bench Verified score is 77.6%. The model is available on OpenRouter at $1 per million input tokens and $4.05 per million output tokens. Hermes or OpenClaw can route Inkling through OpenRouter without additional configuration. These benchmark and pricing figures are presented in the provided material as reported attributes of the model. Both benchmark scores and the OpenRouter pricing are presented without additional context in the supplied material.
The preceding paragraph lists the reported benchmark scores and the model’s access and routing details. This section does not include other benchmark comparisons, implementation guidance, or local-hosting information.
The Inkling model is not designed for running locally, with the explicit statement “You’re not running this locally—not even close.” The model is reported to claim full privacy. In testing described in the provided material, a long 1955-word prompt produced a blank screen rather than a substantive response. A shorter 99-word prompt produced a working game in the same reporting. The reporting also includes the remark ‘”Working” is doing a lot of heavy lifting there.’
This section presents the model’s operational limits, privacy claim, and the observed prompt-test results as reported in the provided material. Direct quotations from the provided material are included where they appear in the reporting. The text above reflects only the items that the provided material reports about operational behavior and prompt testing.
Inkling, the inaugural model from Mira Murati’s Thinking Machines Lab, is presented in the reporting as a Western lab–trained open-source model developed from scratch. The model is described in the material as employing a mixture-of-experts architecture and having achieved notable benchmark results on MCP Atlas and SWE-Bench Verified.
Its trained weights are reported as publicly available on Hugging Face under an Apache 2.0 license. The reporting notes operational characteristics including that the model is not intended for local execution, a claim of full privacy, and sensitivity in prompt handling where a very long prompt produced a blank response while a much shorter prompt produced a working game.
