Inkling 2 Small, a 594-billion-parameter MoE, Available First Through Tinker
Thinking Machines Lab introduced Inkling 2 Small, a multimodal MoE model with 594 billion total parameters and 17 billion active parameters. It is initially available through Tinker, while the company
Thinking Machines Lab unveiled Inkling 2 Small on November 5. It uses a Mixture-of-Experts (MoE) architecture with 594 billion total parameters, activating 17 billion during token processing, while handling text and image inputs together and allowing the level of reasoning effort to be adjusted. Despite its name, the real change in this generation is less the release of a lightweight model than a shift in how large models are introduced into development workflows.
Tinker Opens First
The current access route is Tinker. It can be used for inference in the playground and for sampling and fine-tuning on the platform. The full weights, however, remain closed for now. The company said it would release the weights under the Apache 2.0 license next week after external safety evaluations and red-team reviews are complete, so for now, testing performance in a hosted environment and downloading and operating the model yourself remain separate stages.
What the Performance Means for Coding and Document Work
In its announcement, the company reported that Inkling 2 Small achieved better results than the previous Inkling generation in agentic coding, visual tasks, knowledge work, and reasoning. It particularly highlighted scores on quantitative work evaluations involving financial documents and spreadsheets, as well as long-horizon coding tasks. However, the published comparison table also shows other open-weight models leading in some areas, so a single overall score cannot replace evaluation across every development task.
The Tradeoff Between Scaling Performance and Safety
Long-horizon autonomous work capabilities benefit tasks in which agents chain together multiple tools, but they also increase the risk of trusting results without verification. The model card likewise lists hallucinations, performance degradation in long multi-turn conversations, and variation across languages and fields as limitations. The company advises against using the model for high-risk or safety-critical decisions without additional verification and human oversight. Therefore, the point to watch in this release is less the benchmark ranking itself than the sequence of expanding the deployment scope after completing a safety review before releasing the weights.