inkling vs nemotron 3 ultra vs deepseek v4 pro vs glm 5.2

@thinkymachines by @miramurati recently dropped inkling – their first open-weights model, trained from scratch, full weights on hugging face. key facts:

• mixture-of-experts transformer, 975b total params, 41b active. 1m token context. pretrained on 45t tokens of text, images, audio and video

• controllable thinking effort – you sweep it 0.2-0.99 to trade tokens for quality. they claim it matches nemotron 3 ultra on terminal bench at ~⅓ the tokens

• positioned honestly as "not the strongest model, open or closed"

• benchmarks: hle text-only 29.7, swe-bench verified 77.6, terminal bench 2.1 63.8, aime 2026 97.1, gpqa diamond 87.2. competitive among open weights, below the closed frontier

our test – 3 prompts, single-file html, no libraries, no cdns, no assets. the model designs its own ui – palette, layout, icons, states, everything. we only hand it functional requirements:

1. a calculator – chained ops, divide-by-zero, backspace, %, ±, running expression, full keyboard support

2. a digital clock with a live sweeping second-hand – time, date, weekday, 12h/24h toggle, synced to the wall clock

3. a to-do list – add/toggle/delete/edit, filters, live counter, localStorage that survives a reload

inkling ran at max thinking effort on tinker, the three rivals on @aimlapi

results:

- generation time
#1 nemotron 3 ultra – 2:48
#2 deepseek v4 pro – 7:50
#3 inkling – 13:37
#4 glm 5.2 – 24:54

- lines of code (3 apps, non-empty)
#1 glm 5.2 – 1,701
#2 nemotron 3 ultra – 1,781
#3 inkling – 1,884
#4 deepseek v4 pro – 2,735

the headline: inkling didn't oversell. it lands right in the pack of the open reference models it names in its own release – clearly ahead of nemotron, and trading blows with deepseek and glm. on these (fairly simple) ui-building tasks it's a step behind glm and neck-and-neck with deepseek, exactly where a "solid base model, not a frontier model" should sit

observations:

• design has basically converged. all four shipped genuinely polished uis – glassmorphism, ambient glow, custom svg icons, hover/focus states, keyboard support

• so we stopped screenshotting and started clicking. the calculator. we typed three things a seven-year-old can do:

1 + 2 + 3
• inkling = 6
• glm 5.2 = 6
• nemotron = 7
• deepseek = 4

9 − 4 − 2
• inkling = 3
• glm 5.2 = 3
• nemotron = 2
• deepseek = 7

two "top-tier" open models cannot add three single digits. and the real one, 2 + 3 × 4, which should be 14: nobody got 14. not one of four implements operator precedence. inkling and glm at least return a consistent 20 (left-to-right, the iphone answer). nemotron returns 22, deepseek returns 8: wrong under any interpretation

so we read the code

the twist: the two that broke it are the two that tried to be clever. nemotron and deepseek each built a correct precedence engine - both would return the right 14 - then corrupted the input before it got there. nemotron's input layer duplicates the left operand (2 + 5 × 4 = 22); deepseek's discards the middle one (2 × 4 = 8)

glm and inkling never attempted precedence: they fold left-to-right on two scalars, nothing to corrupt. the dumb approach is the only one that survives a user

so how did inkling do? the release's claim holds up. its calculator is one of only two in the test that actually does arithmetic, and the design is top-tier across all three apps: glassmorphism, full keyboard, aria labels, a to-do that survives reload. it lands exactly where a "solid base model, not a frontier model" should: clearly ahead of nemotron, trading blows with glm and deepseek, and on the right side of the one line that mattered