Thinking Machines Releases Inkling 2 Small: A 594B-Parameter MoE Model

Thinking Machines Lab has released Inkling 2 Small, a large model built on a mixture-of-experts (MoE) architecture. It has 594B total parameters but activates only 17B per inference, scoring 31 on the Artificial Analysis Intelligence Index. More importantly, the model is available for running through the Tinker platform and is planned to be released under the Apache 2.0 license.

594B Total Parameters, 17B Activated: The Efficiency Trade-off of MoE

The most striking numbers for Inkling 2 Small are the combination of 594B total parameters and 17B activated parameters. The core logic of the MoE architecture is to split the model into multiple expert networks and invoke only a subset of them in each forward pass, thereby controlling inference costs while retaining large-capacity knowledge storage. An activation of 17B means the compute cost of a single inference is close to that of a medium-sized dense model, while the 594B total parameters give the model greater knowledge capacity and expressive power.

This design has become increasingly common in the large-model race over the past two years, but the 594B/17B ratio is still noteworthy. The activation ratio is roughly one-thirtieth of the total parameters, indicating that Thinking Machines has pursued fairly aggressive sparsification in expert routing. The benefit of sparsification is cheaper inference; the cost is that routing mechanisms become harder to design and training stability is harder to guarantee. Whether Inkling 2 Small can strike a balance between the two will require real-world usage feedback to judge.

How the MoE (Mixture-of-Experts) Architecture Works

A mixture-of-experts model consists of multiple parallel feed-forward network submodules (the "experts"). Each token is routed at each layer to only a few experts for computation, while the remaining experts do not participate in the forward pass. As a result, the model can have a far larger total parameter count than a dense model, while the floating-point operations per inference depend only on the size of the activated experts. Routing is typically implemented by a lightweight gating network that outputs a weight distribution over experts for each token.

The costs of MoE include: load-balancing losses during training to prevent expert "collapse," memory that must hold all expert weights (even when not activated), and gradient estimation noise caused by the discrete choices of routing. A common industry practice is to pair it with expert parallelism and capacity factor control, which also explains why a ratio of very large total parameters to very small activated parameters is not free to obtain in engineering terms.

What a Score of 31 Means

According to a report by MarkTechPost, Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index. This index aggregates multiple benchmarks to compare the overall intelligence level of different models across the board. A score of 31 falls within a competitive range in the current model landscape, but a single composite score makes it difficult to judge the model's performance on specific tasks.

For developers, what is more informative is the model's actual performance in specific scenarios such as coding, reasoning, and long context. The source material does not provide data on these dimensions, so at this stage it is inappropriate to draw overly specific conclusions about the capability boundaries of Inkling 2 Small. The composite index can serve as an entry-level reference, but it cannot replace targeted evaluations.

What the Artificial Analysis Intelligence Index Is

This index is maintained by the independent evaluation organization Artificial Analysis. It normalizes scores from multiple mainstream benchmarks (covering dimensions such as general reasoning, knowledge, and coding) and aggregates them with weights into a score on a 0–100 scale, used to compare models of different vendors and sizes under a consistent standard.

It is an "aggregate" metric. Its advantages are cross-model comparability and timely updates; its disadvantage is that it masks task-level differences and has limited discriminative power for specific domains (such as code generation or long-document understanding). In addition, scores fluctuate slightly with subsequent benchmark version adjustments and evaluation configuration changes, so it is usually treated as a rough coordinate for a model's capability tier rather than a decisive basis for selection.

The Tinker Platform and Apache 2.0: Lowering the Barrier to Use

Inkling 2 Small runs through Tinker, the model service platform provided by Thinking Machines. Launching the model directly on its own platform indicates that the team wants to connect training, deployment, and tuning into a single pipeline, reducing developers' friction with infrastructure.

The plan to release the weights under the Apache 2.0 license is the most far-reaching piece of information in this release. Apache 2.0 permits commercial use, modification, and redistribution, which is a fairly permissive licensing condition for teams that want to build on the model for secondary development or private deployment. If a model of the 594B class can truly be open-sourced, it would significantly change the accessibility of medium-to-large models—previously, models of this scale could mostly only be used through API calls. The specific timeline for releasing the weights and the full terms are currently mentioned only as a "plan" in the source material, and further official clarification is needed.

What the Apache 2.0 License Actually Means

Apache 2.0 is an OSI-approved open-source license that allows users to freely use, modify, and distribute the software, including for commercial purposes and closed-source redistribution. It requires retaining copyright and license notices, indicating modifications, and provides an explicit patent grant clause—contributors grant users patent licenses related to their contributions, which is particularly important for downstream commercial users in the context of large-model weights.

It should be noted that the license covers the redistribution and use terms of the "weight files" and does not automatically resolve compliance issues in training data, nor does it mean that trademarks or model names can be used freely. For teams seeking private deployment, Apache 2.0 essentially eliminates the legal uncertainty of "separate authorization required for commercial use," but they still need to independently assess liability for model outputs in specific business scenarios.

Potential Impact on the Open-Source Model Landscape

Inkling 2 Small combines two trends: the MoE sparsification approach and open weights. On one hand, 594B total parameters with 17B activated represents an engineering philosophy of "large capacity, low inference cost"; on the other hand, the Apache 2.0 license gives this capacity a chance to reach the community.

If the weights are released as planned, the biggest beneficiaries will be teams that need domain fine-tuning and data privacy. They can run the model locally or on a private cloud without sending data to a third-party API. Of course, the storage and deployment costs of 594B parameters remain substantial; 17B activation only reduces the compute requirements of a single inference and does not turn the model into a lightweight contender. In practice, hardware budgets remain an unavoidable threshold.

The information density of Thinking Machines' release is considerable, but the public details are limited. Key specifications such as the model's specific architecture parameters, training data composition, and context window length have not yet been disclosed. For readers following the direction of open-source large models, Inkling 2 Small is worth tracking, especially community evaluations once the weights are officially released.

One-line Summary

Thinking Machines releases the MoE model Inkling 2 Small, with 594B total parameters and 17B activated, planning to release weights under Apache 2.0.

Paragraph Summary

Thinking Machines Lab has released Inkling 2 Small, built on a mixture-of-experts architecture with 594B total parameters and only 17B activated per inference—an activation ratio of roughly one-thirtieth of total parameters, reflecting an aggressive sparsification trade-off: inference cost close to a medium-sized dense model while retaining greater knowledge capacity. The model scores 31 on the Artificial Analysis Intelligence Index, placing it in a competitive range, but performance on specific tasks has not been disclosed. The model debuts on its own Tinker platform and is planned for release under the Apache 2.0 license, which, if realized, would significantly lower the barrier to private deployment and secondary development of medium-to-large models. However, the storage and deployment costs of the 594B scale remain substantial, and key specifications such as architecture parameters and context length are still to be announced.