Thinking Machines Lab added downloadable Inkling 2 Small weights on Hugging Face on November 12. This is a material development after the model's initial November 5 introduction through Tinker, which SIExpo already covered. We are not presenting that announcement as another model launch: the new event is the availability of published weights for teams considering their own deployment.

The model card lists Apache 2.0 and a mixture-of-experts architecture with 594 billion total parameters and 17 billion active parameters. The model accepts text and images and produces text. “Active” describes the portion used during processing, rather than the size of the complete weight set. The 17 billion figure therefore does not imply the memory footprint of an ordinary dense model with that parameter count.

Practical compatibility needs a separate reading. The checked Transformers main-branch documentation distinguishes the mixed NVFP4/MXFP8 checkpoint from the separate MXFP8 version and states that this integration currently supports only the MXFP8 checkpoint. It describes eager attention because of the model's attention characteristics. This is the state of a particular documentation branch at verification, not a promise about every installed stable release or every inference engine.

The change matters to teams seeking control over their environment, modifications and input processing. Running a model themselves also places infrastructure, updates, monitoring and access-control decisions with the operator. Downloadable weights are not a ready-made managed service. Similarly, the model's licence should not be expanded into a claim that all training data are published or that every future application has already received legal clearance.

SIExpo's editorial interpretation separates three checks. First, does the exact software version support the chosen checkpoint? Second, can the available memory and compute configuration serve the intended workload at acceptable latency and concurrency? Third, is output quality adequate on representative examples from that workload? Successfully loading a model addresses only part of the first question. It does not independently answer the capacity or quality questions that determine whether an application is usable.

For a multimodal application, evaluation examples should include the actual image types and answer constraints the system will encounter. An internal assessment can distinguish missed details, unsupported statements and correctly expressed uncertainty. If the application also triggers tools, permission to act deserves a separate check. Open weights do not establish the safety of an entire agent workflow. These are suggested deployment criteria, not results of tests conducted by SIExpo or additional benchmark claims attributed to the developer.

The practical news is therefore greater choice together with greater operating responsibility. Published weights make a different deployment arrangement possible; they do not establish inexpensive operation on a small computer, a particular throughput level or superiority on every task. A useful decision requires a matching software version, checkpoint, resource plan and evaluated workload. Those details determine what the newly available weights mean for an individual team, beyond the fact that a download now exists.

Editorial verification: 2026-11-15T17:59:15.509180+00:00.