Inkling brings a 975-billion-parameter model with open weights

On paper, Inkling is a heavyweight: Thinking Machines Lab has unveiled its first large language model and released its weights. With 975 billion parameters, it is the largest open-weight model outside China mentioned in the source. The US start-up was founded by Mira Murati, OpenAI's former chief technology officer.

However, making the weights available does not mean the model will run on every computer. In the compact NVFP4 quantization, Inkling requires around 600 GB of RAM for its parameters alone, and as much as 2 TB in bfloat16 format. That puts local operation out of reach for many users. The weights are available, but the required hardware is not automatically available as well.

Technically, Inkling uses an MoE system, meaning many specialized submodels, only some of which perform computations at any given time. Of 256 experts, six are active for each prediction; two shared experts always work alongside them. This means 41 billion parameters are used in each calculation. Text, image and audio processing are also included, as is a context length of one million tokens.

Developers have two specific options: Inkling can be further trained on their own data on the Tinker platform, and a smaller model handles Multi-Token Prediction. It predicts additional tokens that the larger model can then check. According to Thinking Machines Lab, this could speed up output and reduce the average number of parameters required. Whether this calculation holds up has not yet been independently confirmed.

So what does this change in concrete terms? Teams gain a very large, open-weight US model that they can study and adapt with their own data, without being limited to a closed model. However, running it on-site requires a correspondingly large memory and compute environment. For now, the progress lies primarily in access to the weights and in the model architecture—not in straightforward deployment on one's own computer.