thinking machines' inkling-small fixed more bugs than inkling – at 3.5x fewer parameters and 11x cheaper

thinking machines' inkling-small fixed more bugs than inkling – at 3.5x fewer parameters and 11x cheaper

@thinkymachines by @miramurati shipped inkling-small jul 30: 276b total, 12b active, against inkling's 975b/41b. their card already claims small beats its parent on agentic coding – swebench verified 80.2% vs 77.6%

we wanted to see it on code neither model had seen

the setup: @NousResearch's hermes agent cli by @Teknium1, driven through a pty so the /goal ralph loop could run – a judge grades every turn and re-prompts until satisfied. 20-turn budget, file/terminal/browser tools, delegation off so neither could hand work to another model

the apps – two single-file html pages we wrote, 4 visible defects each:

1. tiny town: buildings land mirrored, far ones paint over near ones, autosave fires before the mutation, windows glow at noon.

2. orrery: worlds drawn inside their own rings, speed slider inverted, clicking a planet opens the next one's card, pause stops nothing.

neither model was told how many bugs existed – only what correct behaviour looks like

- bugs fixed
#1 inkling-small – 7/8
#2 inkling – 6/8

- wall clock
#1 inkling-small – 4m 41s
#2 inkling – 32m 57s

- tokens
#1 inkling-small – 48k
#2 inkling – 722k

- cost
#1 inkling-small – $0.10
#2 inkling – $1.08

- tool calls
#1 inkling-small – 56
#2 inkling – 115

observations:

• inkling's town run is the story. it found the mirrored-placement bug small missed – then fixed it wrong, swapping the axes correctly but also changing math.floor to math.round. we replayed all 100 plots clicking dead centre: 75 land on the wrong plot. it reported "re-checked in the running page, togrid – true"

• both verify, but only what they already suspect. each spun up a local http.server and drove the page through the browser console. neither asked "does a building land where i clicked" – the most visible defect in the app. small never formed the hypothesis; inkling formed it, broke it, and confirmed its own break

• 98.9% of inkling's orrery run was input – 626,890 in against 7,023 out, 27m53s, 69 tool calls, to change 7 lines. small changed the same 4 bugs in 2m03s and 19k tokens. 33x the tokens for the same 4/4

• the judge never caught the missing bug. it bounced small three times, all on formatting, and said nothing about the fourth defect sitting unfixed. a ralph loop grades the answer against the goal text, not against the software. every score above is ours, by hand, against a key neither model saw

conclusion: at a quarter the parameters, inkling-small fixed one more defect than its parent on a fifth of the money and a fifth of the clock!