Thinking Machines Releases Inkling, an Open-Weights 975B Parameter LLM

Inkling is a 975B parameter open-weights language model released by Thinking Machines AI. At that scale, it sits alongside the largest publicly accessible models available today.

The open-weights designation matters more than the parameter count. Engineers can download, inspect, fine-tune, and self-host the model without negotiating API access or accepting usage restrictions that closed providers impose. For teams building on top of foundation models, that changes the economics and the control surface substantially.

At 975B parameters, Inkling operates in a range where inference demands serious hardware. Expect multi-GPU or distributed setups to run the full model at reasonable throughput. Quantized versions, if the team ships them, would reduce that barrier, but at this scale even aggressive quantization leaves a significant memory footprint. Builders evaluating Inkling need to plan infrastructure before committing to it in a production path.

The release is notable regardless of benchmark position. Most open-weights releases cluster well below this scale. MoE architectures have let some models punch above their active-parameter weight, but a dense or near-dense model at 975B is a different infrastructure proposition entirely. The team has not published detailed architecture specifics in what is publicly visible from the announcement, so the precise design remains to be confirmed from the release materials.

For solo founders and small teams, Inkling is not a drop-in replacement for managed API inference unless cloud providers pick it up. The practical path is either a managed fine-tuning provider that adds support, or an in-house cluster. Neither is trivial.

What shifts here is the frontier of what is openly licensed and inspectable. Each release at this scale narrows the gap between closed and open capability, and that gap is where most of the practical leverage in AI product development currently lives.

Source