We noticed that Inkling 2 Small activates only 17 billion parameters per computation while totaling 594 billion, and it scored 31 on the Artificial Analysis Intelligence Index. What matters is not merely the leaderboard improvement, but that the model will initially run under closed weights, complete external safety testing, and only then open all weights.
What this is
Inkling 2 Small is the team’s second-generation model. It uses a Mixture of Experts architecture (MoX, a design that invokes only part of the model’s parameters at a time to control computational cost), supports image inputs, and offers adjustable reasoning effort. The team says it outperforms the previous generation in agentic coding, vision, knowledge work, and reasoning tests, with performance improved through more efficient pretraining and unified reinforcement learning training.
The model will initially be available only through Tinker for inference and fine-tuning; full checkpoints are planned for download when the weights open. Under the official schedule, the model will be released under the Apache 2.0 license after completing external safety testing and red-team testing.
Industry view
We believe this “hosted first, open later” release model could become a practical way for more capable models to balance controllable risk. The closed phase allows access and costs to be managed, while expanding the ecosystem after testing gives developers and enterprises time to evaluate the model in advance.
But the open-weight timetable remains a promise, not a delivered fact. Safety testing may expose new vulnerabilities, and a leaderboard alone cannot fully verify a model’s scale or training process. Another risk is that activating 17 billion parameters does not automatically mean a low deployment barrier; GPU memory, inference frameworks, and serving stability will still determine whether enterprises can adopt it. The team also says performance exceeds the previous generation, but has not provided complete comparative details for each test.
Impact on regular people
For enterprise IT: Once the weights open, organizations will be able to download checkpoints and fine-tune the model. In the short term, it is better suited to validation by organizations with compute resources and engineering teams than to direct replacement of existing systems.
For individual professionals: If its image understanding, coding, and knowledge-work capabilities become reliably usable, they will gradually enter research, document processing, and software collaboration workflows, though results will still depend on the specific task and integration method.
For the consumer market: The closed-service phase means ordinary users can initially try it only in Tinker Playground. Whether open weights lead to cheaper, more localized products will depend on whether the supporting toolchain matures.