Are "small reasoning models" the next big shift? What should we actually be measuring?

For a model running locally on a fairly narrow task, how much general knowledge do we actually need, and how much reasoning capability could we get without it ? SRMs are interesting for obvious reasons, but I went down this rabbit hole after listening to Ben Lorica's (advisor at Databricks) chat with Zuzanna Stamirowska from Pathway (BDH). Ben keeps coming back to this broader theme of how "specialized AI is getting easier to build" and the Kumo RFM angle, but it opened an interesting thread around small reasoning models. Pathway’s ARC-AGI-1 result makes an interesting case for small models hitting the cost-accuracy Pareto frontier. The premise is that if a model is built to reason natively in its latent space, it might not need billions of parameters absorbing Reddit and Wikipedia just to solve logic puzzles. They described a use-case of long-horizon reasoning within a bounded domain as a target (like security investigations, tickets analysis – a real case I know from a major bank, etc.) It's obvious that just because a large model does well in 20 languages. I don't need that for work tasks. There is definitely a market for compact models with substantial reasoning ability. Also because architectures like BDH handle state and memory differently than standard transformers, the pitch is that they avoid catastrophic forgetting (learning continuously from new examples at inference time without wiping past skills). The question is how to evaluate this without getting lost in marketing claims. Here is how I'd break it down • Compactness: Low parameter count, but what are the actual inference memory and compute requirements? • Few-shot adaptation: Does it adapt through context or actual parameter updates? • Training efficiency: How much data did it actually need to pick up the underlying capability? • Continual learning: Does post-deployment experience produce persistent improvements without degrading earlier skills? For people running small models locally: what workload expose the difference between a compact model that just follows in-context examples versus one that actually learns reusable rules?