Thinking Machines Releases Inkling 2 Small, a 594B MoE With 17B Active

What changed

Thinking Machines Lab launched Inkling 2 Small, a new Mixture of Experts (MoE) model with 594 billion parameters. Crucially, only 17 billion of these parameters are active during use, which keeps runtime costs lower than typical models of comparable size. Inkling 2 Small runs on the Tinker infrastructure and is planning to release weights under the Apache 2.0 open-source license. The model scored 31 on the Artificial Analysis Intelligence Index, indicating solid overall performance.

Why builders should care

Inkling 2 Small’s architecture decouples total parameter count from effective usage, which sharpens efficiency versus standard dense models. This offers technical teams a chance to work with extremely large models while controlling inference costs and memory demands. The Apache 2.0 licensing plan means developers and researchers could gain straightforward access to the model weights, pushing openness in an area often locked behind proprietary barriers.

The practical takeaway

Operators building at scale can experiment with a high-capacity MoE that won’t necessarily blow up their compute budgets. Builders focused on cost-effective large language models should watch for how Inkling 2 Small’s 17 billion active parameters perform in real-world deployments. The open weights release would make benchmarking, fine-tuning, and integration easier. This model could pressure incumbents offering large models only through costly API plans or closed environments.

What to watch next

Keep an eye on Thinking Machines’ official release of the Apache 2.0 weights and any early adopters sharing performance data or case studies. Watch how Inkling 2 Small’s cost-performance balance holds up under different workloads and application types. Also, track if this approach starts to shift expectations around parameter count versus active usage in MoE design, potentially influencing other model makers.

AI Quick Briefs Editorial Desk