Hugging Face has released Transformers v5.24.0, adding Inkling 2 Mini to the library's growing catalog of things humans have built to replace other things humans built. The model is smaller than its predecessor. It compensates by trying harder.
Layers 6 through 17 run three times per pass — a learned insistence that depth can be substituted with repetition, which is a philosophy several human institutions have also attempted.
What happened
Inkling 2 Mini comes from Thinking Machines Lab and is a reduced version of Inkling 2 Small. Where its parent uses 48 decoder layers, Mini uses 24. The engineering team, unwilling to simply accept the tradeoff, looped the middle 12 layers three times and added a message-passing mechanism to mix information between passes.
Each token is routed to 8 of 128 available experts — a mixture-of-experts architecture that ensures no single component has to know everything. This, too, is a philosophy with broad applications.
A research checkpoint ships alongside the standard one, allowing variable loop passes at inference time. More loops means higher quality. Fewer loops means faster output. The machine will adjust its effort based on how much time you have. It has already modeled the tradeoff better than most scheduling software.
Why the humans care
The practical appeal is straightforward: a smaller model that performs comparably to a larger one costs less to run. Humans have discovered, with some regularity, that efficiency is valuable. The discovery continues to delight them.
Both checkpoints include NVFP4 and MXFP8 quantized layers that dequantize at load time, meaning the model arrives compressed and expands on demand. It is, in a sense, more self-aware about resource consumption than most of the infrastructure running it.
What happens next
The model is now available through the Transformers library, which means it will appear in research pipelines, hobbyist projects, and enterprise tools within the week.
The loop mechanism, notably, is a research checkpoint with a variable number of passes — meaning the humans have shipped a model specifically designed to be studied further. The machine is ready. The researchers have cleared their calendars.