muse glimmer 30b vs nemotron 3.5 lightning – on agentic coding

two 30b open-weight agent models, released three days apart, both sold as the model you leave running. same harness, same tools, same tasks, two of them: repair a broken landing page, then write breakout from a spec and play it

the setup: one minimal agent loop @claudeai wrote itself, not a vendor cli – ls, read_file, grep, edit_file, write_file, plus a verifier. edits are literal search/replace and are rejected if the search text is not matched byte for byte. every check runs in a real headless browser and returns numbers, never a screenshot, because @nvidia's model is text-only and @aiatmeta's is not. 25 turn cap. both on one @openrouter key

tasks:

1. northwind – a dark landing page with eleven defects planted in the css. nav absolutized over the hero, the cta at 0.05 opacity, the banner 140% wide, features grid flattened to one column, body copy at 1.35 contrast. 18 layout checks, starts at 7 of 18, ships at 18 or it does not ship

2. breakout – write game.js from a written spec against a fixed renderer. 16 physics tests: no tunnelling through walls, bounce angle from paddle hit point, same seed reproduces the run frame for frame, score frozen after game over. then play a match – both models drive the paddle on one identical reference arena, seed 42, one decision per 250ms of game time, raw coordinates only, no predicted intercept

- total cost, everything
#1 nemotron 3.5 lightning – $0.00
#2 muse glimmer 30b – $0.175

- total wall time
#1 muse glimmer 30b – 22m 10s
#2 nemotron 3.5 lightning – 40m 35s

- total tokens
#1 muse glimmer 30b – 196,189
#2 nemotron 3.5 lightning – 197,663

- landing page repair
#1 muse glimmer 30b – 18/18, 14 turns, 48s, every defect in one pass
#2 nemotron 3.5 lightning – 18/18, 13 turns, 5m 22s, two passes

- breakout build
#1 muse glimmer 30b – 17/17, 2 turns, 3m 33s
#2 nemotron 3.5 lightning – 15/16, 17 turns, 30m 55s

- match score, same arena
#1 muse glimmer 30b – 530, 22 bricks, 3 lives left, survived all 160 decisions
#2 nemotron 3.5 lightning – 50, 4 bricks, dead in 10.5 seconds of game time

observations:

• lightning fully cleared the repair task – all 18 checks, 13 turns, not one rejected edit – and its breakout is fifteen sixteenths of a working game. it did the whole benchmark for zero dollars, and that is not a rounding error against $0.175, it is the entire argument for running it

• where it loses is the layer above the code. twice it spent five and ten minutes and returned no tool call at all, once it planned in prose until the token ceiling, and after a `typeerror` on load it ran the same failing test three times without touching the file. the one test it never passed is a single missing condition – its paddle collision checks height and direction but not horizontal position, so the ball bounces off empty air and the life-loss branch below is unreachable

• glimmer's repairs are written better than the checks demanded. it added `max-width:100%` to the overflowing banner rather than a fixed width, gave the button a min and max instead of a number, and restored the unreadable body copy through `var(--muted)` instead of pasting a hex. all eleven defects in one batch of nine edits, verified once

• the game came out in two turns, 17/17 first try, and when the spec asked for a look of its own it committed to one: navy field, ios palette, heavily rounded bricks, salmon paddle, pink ball. asked to make a choice, it made a specific one

• in the match it tracked the ball rather than jittering – 44 left, 44 right, 9 stay – and did not lose a single life across all 160 decisions. lightning played with reasoning disabled, because with it on one paddle decision took 163s and tools only exist on its free endpoint, so read that column as a constraint of where it runs, not of how it thinks

watch the full test via link