Commentary on X
Inkling small is ~3.5x smaller but the performance looks insane? I think the comparison table here gives quite abit of insight as to what requires larger model capacity (things like TerminalBench and SimpleQA) while "reasoning" can be condensed to some "core capability"
4:54 PM · Jul 15, 2026@finbarrtimbers Yep probably the coolest part? Plus, IFBench being the one where Inkling small does better further implies that lexical/constraint type IF evals probably have not much real-world use anymore
5:06 PM · Jul 15, 2026Capabilities wise, the model looks to be more general purpose rather than the standard pure coding agent type models. Makes a ton of sense for their business model via Tinker and sounds very much like Cohere. Design-wise, the font and color choices scream Claude to me
4:54 PM · Jul 15, 2026View all
> adjusting per-token cost Interesting. I guess this means adjusting length penalty based on thinking effort? This likely directly caused the compressed language used in cot.
4:54 PM · Jul 15, 2026