On November 20, Hugging Face introduced its integration of Thinking Machines Lab’s Inkling 2 Mini, a multimodal mixture-of-experts model with roughly 26 billion total and three billion active parameters per token. Its middle layers run three times while sharing weights. Announced inference options include Transformers, llama.cpp, MLX and SGLang. Its stated maximum context is one million tokens, which does not establish that every computer can use that input comfortably.

Memory and time answer different questions

The practical interest is the tradeoff between stored weights and executed computation. Reusing a component saves storage compared with keeping a separate component for each step. Each repeated computation still has to run. Model size alone therefore cannot justify buying a machine or promising a delivery time to a customer.

The model card remains necessary when selecting the precise variant, format and usage conditions. Multiple inference options also require an operational choice: one team may value a desktop application while another needs a managed server. Either way, record the actual software versions rather than attributing every observed difference to the model’s name.

How we would evaluate a configuration

Our proposed pilot would start with one actual task, such as extracting information from several product descriptions. Keep the input fixed and separately record time to first response, time to completion, memory consumption and factual mistakes. SIExpo has not conducted this experiment and reports no original speed measurement here.

Next, we would add an image and a longer document, watching where the setup becomes inconvenient. An explicit lack of an answer also matters: insufficient evidence should be acknowledged. An apparently fast response that requires extensive human repair may cost more than a slower, verifiable result.

For an agent using tools, we would initially allow read-only access to a copy of the data. Writing, sending messages and making payments require separate permissions. Local execution may help a team control information, but the computer’s physical location does not prove that every connected tool avoids external services.

We would also test recovery. Interrupting a run should leave a record of which steps finished and which remain uncertain. Repeating the entire workflow without that record could duplicate a later action even when the language model itself runs locally.

The release adds a concrete option for local experimentation. Whether it fits a particular job depends on the complete workflow, including human revision and recovery after an error, rather than on the size of its downloadable weights alone.