Inkling is an open AI for text, image and sound

Thinking Machines Lab has released Inkling, an AI model that can receive text, images and sound and leave text in response. The model weights, i.e. the learned values ​​that control the model, are available on Hugging Face under the Apache 2.0 license. The release is primarily aimed at developers and researchers who want to build or adapt their own AI services.

Inkling is multimodal. This means that the same model can work with several types of substrate. For example, a development team can combine an image with a question or have the model interpret an audio file. Third-party audio support was still a work in progress when Hugging Face published its guide.

The model has 975 billion parameters in total, but uses 41 billion at a time. That technique is called a blend of experts: only certain parts of the model are activated for each step. It reduces the computation compared to using the whole model at once, but does not make Inkling small.

Hugging Face states that the full version requires about 2 terabytes of graphics memory and that the compressed NVFP4 version requires about 600 gigabytes. It's server level, not a regular laptop. Less compressed variants are available from other players, but they are not the same as the version Thinking Machines Lab evaluated.

Thinking Machines Lab reports its own comparisons against other models. At the same time, the company writes that several results were produced with a different model checkpoint from the one released. The figures should therefore be read as the supplier's tests, not as independent proof that Inkling is better than other models.

The model card warns of incorrect answers, uneven quality between languages and risks in open installations. Developers are advised to test the quality, safety and fairness of their own use and put protection around the model rather than relying on its built-in rejections.

Därför spelar det roll

Open model weights give Swedish developers greater opportunity to review, adapt and run a model in their own infrastructure. At the same time, Inkling shows that open access does not mean simple local operation. The cost of hardware, testing and security can be greater than the actual access to the model.

Det här kan du göra

1. Read the model card, license and hardware requirements before planning a test.
2. Start with a limited task and your own non-sensitive test material in text, image or sound.
3. Compare the result with the original source and test Swedish, security and cost in the environment where the model will be used.