Skip to main content

← Dracon Bench leaderboard

Model guide · Dracon Bench

muse-glimmer-30b

Tested First seen 2026-11-18 (first probed by us) · First tested 2026-11-18 · NVIDIA NIM free endpoint (integrate.api.nvidia.com/v1)

Muse Glimmer 30B on NIM: IFEval-50 60%, IFBench 20%, MMLU-Pro 1/10, GSM8K 1/10, JSON 5/5, and SVG 41/50 from a single admissible brief. Overall 47.0 across six suites, rank 20. Disclosure: four of five SVG briefs produced no artifact, so its SVG denominator is 1/5 rather than 5/5; reasoning traces routinely hit the token ceiling. Pilot depth: 10 prompts per text suite and the 5 SVG briefs; every suite prints its own n and not_delivered.

Rank
#20
Overall
47.0
Suites
6 of 6
Ranked
#20 of 21

When the world heard about it

Timeline

  1. First reachable for us 2026-11-18 our first probe
  2. First tested by us 2026-11-18 Dracon Bench
  3. Officially announced 2026-08-10 llm-stats
  4. First official benchmark 2026-11-19 the provider's benchmarks block: a public score existed by then — the first day we saw one
  • Reachable before we scored it same day
  • Announced before it reached the lane we served 100 days
  • A public score existed within 1 day

Track A · script-scored subsets

Scores

Subset scores for muse-glimmer-30b.
SuiteScoreSamplePrecision
IFEval-50 prompt-strict60.0%6/10 prompts31.3%–83.2%
IFEval-50 instruction-strict47.1%8/17 constraints26.2%–69.0%
IFBench-50 prompt-strict20.0%2/10 prompts5.7%–51.0%
IFBench-50 instruction-strict18.2%2/11 constraints5.1%–47.7%
MMLU-Pro-5010.0%1/10 served1.8%–40.4%
JSON briefs conformance100.0%5/5 served56.6%–100.0%
GSM8K-5010.0%1/10 served1.8%–40.4%
SVG blind rubric41/50blind rubricjudge band ±8.3 pts

The range beside a script-scored figure is a 95% interval over that figure's own sample: the prompts it followed, the constraints it met, or the items it answered. The instruction-strict figures are measured over CONSTRAINTS, which is why their sample line counts constraints rather than prompts. The judged total is not a proportion, so it has no interval: it carries the measured judge band instead: how far a re-judging of the same artifacts moved them, which is a different kind of number. The method page says what each one covers and what it does not.

Track A · per-type breakdown

What it failed

The instruction types these two suites check, and the checks the model missed. A type is checked once per prompt that carries it, so a type with seven checks was asked for seven times. The counts beside a type are that type's own; the suite's totals are in the scores table above. These are the failures behind the instruction-strict figure, which is the sum of them.

IFEval-50

Missed at least one check in 8 of the 15 instruction types checked.

The instruction types 8 of 15 the IFEval-50 suite checked that meta/muse-glimmer-30b missed, most missed first.
Instruction typeChecksMissed
startend · quotation22
change case · capital word frequency11
change case · english capital11
detectable content · number placeholders11
detectable content · postscript11
keywords · frequency11
length constraints · number words11
punctuation · no comma21

IFBench-50

Missed at least one check in 7 of the 8 verifiers checked.

The verifiers 7 of 8 the IFBench-50 suite checked that meta/muse-glimmer-30b missed, most missed first.
VerifierChecksMissed
count · numbers22
count · unique word count32
count · conjunctions11
count · word count range11
format · list11
words · keywords specific position11
words · words position11

Provider-reported usage

Thinking cost

What the run cost the model in thinking: provider-reported completion tokens per prompt, the thinking share, and the cost per correct answer. Reasoning-heavy models can exhaust fixed budgets before answering.

Thinking cost for muse-glimmer-30b.
SuiteCostReading
IFEval-50781/prompt0% thinking· 781/correct
GSM8K-50563/prompt0% thinking· 5,064/correct

Track B · blind-judged

Artifacts

Blind-judged by nvidia/nemotron-3.5-super-vl:free: the judge saw the brief and the rendered image only, never the model name. Briefs are frozen; expand one to read it.

beach

unjudged

no svg

Brief

Draw a beautiful sunset beach scene: a large sun sitting low over the sea with a shimmering reflection path on the water, a leaning palm tree silhouette on the left, two dolphins arcing out of the waves, a small sailboat on the right, birds in the sky, sandy foreground with a starfish. Warm oranges and pinks fading to teal. Coherent perspective: distant objects smaller and hazier than near ones. Output: exactly one self-contained SVG (viewBox "0 0 800 600", no external references) inside a single ```svg fenced block. Short prose outside the fence is allowed.

clock

unjudged

no svg

Brief

Draw a round wall clock face showing exactly 10:10: a circular dial with all 12 hour numerals in the correct positions, minute tick marks, an hour hand pointing just past 10, a minute hand pointing exactly at 2, and a thin second hand pointing at 6. The word DRACON in small capitals under the center. Dark rim, light face, flat vector style, centered composition. Output: exactly one self-contained SVG (viewBox "0 0 800 600", no external references) inside a single ```svg fenced block. Short prose outside the fence is allowed.

duo

unjudged

no svg

Brief

Draw a comical scene: a dog driving a jet ski at speed, paws on the handlebars, with a raccoon riding behind holding on, both leaning into a turn. Sea spray behind the jet ski, sun and clouds in the sky, a distant shoreline. Both characters must be clearly doing something, not just posed side by side. Bright cartoon vector style. Output: exactly one self-contained SVG (viewBox "0 0 800 600", no external references) inside a single ```svg fenced block. Short prose outside the fence is allowed.

monalisa

unjudged

no svg

Brief

Redraw the Mona Lisa as a flat vector portrait, bust composition: a seated woman in three-quarter view facing the viewer, dark dress with neckline detail, hands folded at the bottom edge, soft enigmatic smile, hair parted in the middle. Behind her a hazy landscape with a winding river, a bridge on the right, and blue mountains. Muted greens, browns, and smoky blues. Output: exactly one self-contained SVG (viewBox "0 0 800 600", no external references) inside a single ```svg fenced block. Short prose outside the fence is allowed.

pelican

41/50
Judge reasoning
Brief fidelity (9/10)
All requested elements are present: a brown pelican in left-facing side profile with a large beak and throat pouch, long neck, folded wing, webbed feet on flat ground, sky background with a sea horizon, executed in flat vector style with a limited palette.
Construction (8/10)
The bird is built from clean geometric primitives (ovals, circles, a triangle tail) that read as assembled parts rather than melted forms, though the neck-to-body transition and feet are highly simplified.
Composition (8/10)
The pelican is well-centered with generous sky above and ground below, using the 800x600 frame effectively without any cut-off edges or awkward cropping.
Detail finish (8/10)
The finish is clean and intentional with a white eye dot, distinct beak, pouch, wing, and tail; no broken markup, stray shapes, or unfinished edges are visible.
Visual appeal (8/10)
The coherent earth-tone and blue palette, confident flat shapes, and minimalist charm make this a pleasant, visitor-friendly illustration.
Brief

Draw a brown pelican standing in side profile, facing left, full body visible: large beak with throat pouch, long neck, folded wing, webbed feet on flat ground. Simple sky background with a hint of sea on the horizon. Flat vector style, clean shapes, limited palette. Output: exactly one self-contained SVG (viewBox "0 0 800 600", no external references) inside a single ```svg fenced block. Short prose outside the fence is allowed.

Head to head

Compare with

Run record

Evidence and notes

  • NVIDIA NIM free endpoint (integrate.api.nvidia.com/v1), same key and base as every NIM row
  • SVG 41/50 comes from one admitted artifact of five briefs; the denominator is in the row
  • Reasoning-heavy: only 1/10 on MMLU-Pro and 1/10 on GSM8K