GoldieBench deep dive
Inkling's full benchmark breakdown.
A 975B open-weights frontier model — yours to own and run. Every number below is either one of my own judged one-shot builds - playable on this site - or an outside result with its source linked. Nothing display-only, nothing vibes.
01 · The headline numbers
02 · Every benchmark, bar by bar
All 50 scored tasks, best first. Hover for the judge's comment; click through for every model's build on that task. Or overlay other models in the interactive graphs.
03 · Where it wins - with the judge's own words
“Gorgeous dense glyph rain with katakana/symbol mix, glowing trails, and an Orbitron title that reads beautifully; the hue-shifting green-to-blue gradient is striking but drifts slightly from canonical Matrix green, keeping it just shy of the top.”
“A polished 3D Breakout in Three.js with a gorgeous gradient title, glowing rainbow brick wall, paddle/ball follow, trail dots and live score badge — clearly renders and is on-brief. Held back from top spot by the loose 2D collision math on a 3D perspective view (paddle bounce/wall bounds are hardcoded and can feel off) and lack of lives/win state, but visually it's a strong, shippable entry.”
“Renders a vivid, colorful WebGL aurora curtain with clean green/cyan/purple bands, ground plane, and tasteful title overlay — clearly on-brief and polished; but the curtain reads as a contained rectangular slab rather than a sky-spanning flowing veil, and the mountain silhouette is invisible, keeping it just short of the field's best.”
04 · Where it struggles - quoted, not hidden
Boards that hide the weak rows are brochures. These are Inkling's lowest scored builds, verdicts unedited:
“The core plasma effect fails to render — the screenshot shows only a dark background with UI chrome, because the code uses the nonexistent `new THREE.Renderer` instead of `THREE.WebGLRenderer`, so the canvas never appears. The title and controls are polished, but the entire hypnotic field brief is broken.”
“Only the title overlay and hint text render — the entire Three.js scene (black hole, accretion disk, stars, lensing) is absent from the screenshot, so the core brief is completely missing. The gradient headline is polished but the visualisation failed to render.”
“The 3D scene fails to render — the canvas is entirely black with only the UI overlay (title, hint, controls) visible, because the code calls `wallGeo(w,h,d)` as a function when `wallGeo` is a BoxGeometry instance, throwing an error that halts scene setup. Nice atmospheric typography and control styling, but the dungeon crawler itself is broken.”
“The title typography is gorgeous with a clean gradient and neon glow, but the actual particle system barely renders — only a handful of scattered dots and one clustered blob are visible instead of 8000 swirling particles, so the core brief (sculpting a particle swirl) is effectively broken/invisible.”
05 · Category breakdown vs the whole field
Games (23 tasks)
Others (3 tasks)
Pages (3 tasks)
Sims (12 tasks)
Visuals (9 tasks)
06 · Outside signals - every row sourced
My bench measures one thing: judged one-shot builds. These outside rows measure other things - kept separate, never blended into the GoldieBench average, every value linked to where it comes from.
Vendor-published numbers
| Benchmark | Result | Source |
|---|---|---|
| SWE-bench Verified | 77.6% | thinkingmachines.ai |
| SWE-bench Pro Public | 54.3% | thinkingmachines.ai |
| Terminal-Bench 2.1 | 63.8% | thinkingmachines.ai |
| AIME 2026 | 97.1% | thinkingmachines.ai |
| GPQA Diamond | 87.2% | thinkingmachines.ai |
| IFBench (chat) | 79.8% - ahead of every model in its comparison table incl. Fable 5 and GPT-5.6 Sol | thinkingmachines.ai |
Independent evaluations
| Benchmark | Result | Source |
|---|---|---|
| Own framing | explicitly 'not the strongest overall model' - a generalist optimised for customisation, Apache 2.0 weights | techcrunch.com |
From the source guides
| Measure | Result | Source guide |
|---|---|---|
| SWE-bench Verified (agentic coding) | 77.6% — ahead of Nemotron 3 Ultra (70.7) | /ink-machine |
| Terminal-Bench 2.1 (best harness) | 63.8% | /ink-machine |
| GPQA Diamond (reasoning) | 87.2% | /ink-machine |
| Total / active params | 975B / 41B (MoE, open weights) | /ink-machine |
07 · Head-to-head records
08 · Methodology + honest limits
Every GoldieBench score is a one-shot, single-file build from an identical prompt - no retries, no hand-fixing - rendered for real, screenshotted, and scored 0-10 by one judge model on one rubric across the whole field. Failures score as failures. What this bench does NOT measure: multi-turn agent work, long-context recall, or API latency - that is what the sourced outside rows are for. Full method: /methodology.
Frequently asked questions
01How does Inkling perform on GoldieBench?
Inkling averages 6.07/10 across 50 scored one-shot build tasks, ranking #17 of 17 ranked frontier models, with 0 task golds, 0 silvers and 3 bronzes. Every score is a real judged build you can open and play on this site.
02What is Inkling best at?
Its strongest scored build is Matrix at 8.4/10. By category it averages Game 5.9, Other 7.1, Page 7.3, Sim 5.4, Visual 6.5 on the bench.
03How much does Inkling cost?
Inkling is open-weights — a 975B-parameter (41B active) Mixture-of-Experts model whose full weights are public on Hugging Face. You run it on your own key through Tinker's OpenAI-compatible endpoint (usage-based, ~$0.33/M sampling, 50% off at launch), or via Together / Fireworks / Modal / Databricks / Baseten. Benched here one-shot at medium reasoning effort via Tinker.
Source ledger
- 01Official vendor site: thinkingmachines.aithinkingmachines.ai
- 02Own framingtechcrunch.com
- 03SWE-bench Verified (agentic coding)agentos.guide
Every model's breakdown
Run this stack yourself.
Every demo on this bench was built inside the Agent Operating System — one prompt, one shot, single HTML file out. The Agent OS, the prompts, the templates, the weekly walkthroughs and 4,000+ founders shipping with it every day all live inside the AI Profit Boardroom.