Thinking Machines
Inkling
A 975B open-weights frontier model — yours to own and run.
Official vendor source
Inkling is built by Thinking Machines — see the vendor's own product page, pricing, and docs at thinkingmachines.ai.
Reference benchmarks for Inkling
These are external benchmarks I pulled from the source comparison guides on agentos.guide — SWE-bench Verified, DRACO, Kilo plan rubric, build-time measurements, vendor-reported coding scores. They are not goldiebench medal scores (those come only from same-prompt one-shot creative coding tasks in the matrix). I surface them here so the spec sheet for Inkling is honest about what's measured.
What is Inkling?
Inkling is the Thinking Machines frontier model with a 1,000,000 tokens context window, released 2026-07. Tagline: A 975B open-weights frontier model — yours to own and run.. Official source: thinkingmachines.ai.
Pricing detail. 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.
How I use it inside the Agent OS. Benched on GoldieBench one-shot through Tinker's OpenAI-compatible endpoint at medium reasoning effort, then headless-playtested on the same rubric as the whole field. In the Agent OS it's wired into the opencode tab on your own Tinker key — the Ink Machine.
What I built with Inkling
Every model on Goldie Bench gets the same fixed prompt set — one shot, single HTML file out — and I score the result 0–10 inside the Agent Operating System. Here's what Inkling shipped on the bench: 50 one-shot demos across 1,000,000 tokens of context. Of those, 50 are scored against the field with my honest 0–10 from the source guides at agentos.guide.
Strengths
- Genuinely open-weights — the full 975B model is public on Hugging Face; run it on your own key, no black box
- Best one-shot builds are 2D / animation / web — a matrix-rain that topped its task (8.4), plus arcade, fractal, aurora and a mini web-OS all judged shippable (7.6–8.2)
- Frontier-class agentic coding for an open model — 77.6% SWE-bench Verified, ahead of Nemotron 3 Ultra
- 1M-token context, native multimodal (text/image/audio), and a controllable thinking-effort dial
Trade-offs
- One-shot 3D games are weak — three.js dungeons/racers render a title screen but no playable scene, like most open models (crypt 2.5)
- Physics and particle sims are hit-or-miss — black-hole, plasma and cloth one-shots often render dark or static (2.3–3.5)
- Not the strongest overall — the closed frontier (Fable 5) still tops the raw benchmarks; Inkling trades peak for ownership
Best for
- Owning a frontier model instead of renting one — on your own key, pennies per build
- Generative visuals, data-viz and single-file web builds you want one-shot
- A customizable open base you can fine-tune on Tinker for your own domain
Every demo by Inkling
50 live demos, sorted by category. Click any tile to play the actual one-shot result. Verdicts and 0–10 scores are pulled from the source guides where I posted them publicly.
▶ LIVECrypt
Game
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.
▶ LIVEDragonrealm
Game
Renders a clean, atmospheric snowy scene with polished Cinzel title, mountains, and vignette, but the snow particles are oversized blocky squares (size bug from the broken forEach loop) that read as debris, and crucially the sword isn't visible in-frame — undercutting the core 'draw your sword' brief. Solid mood, but generic and missing key interactivity payoff.
▶ LIVEDoom
Game
Renders a working raycaster corridor with title, HUD and crosshair, but monsters are just flat red rectangles (one dominating the view here) with no sprite art, and the all-red palette makes walls/enemies indistinguishable. Functional but visually crude and generic compared to the field's best.
▶ LIVERacing
Game
Strong neon title/HUD polish and the car mesh renders cleanly, but the track is essentially invisible in the dark scene — it reads as floating boxes in space rather than a road, undermining the core 'track and obstacles' brief. Functional third-person camera and physics in code, but the visual presentation looks broken/unclear compared to the field's best.
▶ LIVESkyrim
Game
Renders a clean low-poly 3D scene with trees, rocks, and a ruin pillar plus polished title/HUD, but the world feels sparse and floaty with no visible ground plane, no mountains despite the promise, and generic pointer-lock exploration that falls well short of the field's best.
▶ LIVETwilightvale
Game
Clean UI and title, but the 3D scene barely renders — the terrain/ground plane appears invisible or floated away, leaving just a floating sphere and stray tree trunks against a bare background, so it reads broken rather than an open-world realm. Nice HUD polish and combat-log styling can't rescue the failed terrain/world presentation.
▶ LIVEVoxelcraft
Game
Renders a clean, colorful voxel terrain with nice shadows, gradient title, and polished HUD; but the flat pastel palette looks more candy than Minecraft, placement only works on y=0 ground plane, and the pointermove-based 'orbit' fights every click since dragging is never gated — solid but generic and mechanically shaky.
▶ LIVEGtafoot
Game
Renders a clean neon 3D city with roads, cover barriers, pedestrians and a wanted-star HUD, but the giant title/subtitle/DOM cover-barrier overlay dominates the frame and the scene reads generic voxel blocks rather than a polished GTA-style on-foot game; functional but not a task winner.
▶ LIVEGtadrive
Game
Renders cleanly with a polished HUD, wanted stars, minimap, and 3D box cars with headlight glow, but the ground reads as a flat dark void with barely visible grid lines and no real city/buildings, making it feel sparse rather than a living open city. Competent and functional but generic-to-mid versus the field's best.
▶ LIVEAipbpromo
Page
Renders cleanly with a striking gradient logo, glowing 3D torus and polished glassmorphic stat cards, but it's a single static-ish scene—no multi-scene 'Remotion-style' sequencing, no count-up stat animation, and the ring visually clashes with the center CTA, falling short of the cinematic auto-playing video brief.
▶ LIVEParachute
Game
Renders cleanly with cohesive jungle scene, glowing target ring, jumper, and parachute plus a working land-in-clearing state; but the chute sits detached above the figure (ropes disconnected) and the flat green-monochrome palette reads generic rather than winning.
▶ LIVEFlightsim
Game
HUD panel is polished and on-brand, but the screenshot shows a broken render — a giant close-up brown cross (the plane fills the frame) against a black void with no visible terrain, runway, sky, or clouds; the camera/scale is clearly wrong and the sim reads as non-functional.
▶ LIVEArcade
Game
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.
▶ LIVEDogfight
Game
Clean neon HUD, functional core loop (mouse-follow, shooting, spawns, collisions/lives), but visually thin — the player 'plane' reads as a flat pink block and clouds are drab boxes with no dogfight drama or enemies visible in frame. Works and is polished at the UI level but generic and low-fidelity for the 3D combat feel.
▶ LIVEDragonflight
Game
Strong HUD styling — title, Speed/Rings stats, and a polished fury meter all render crisply — but the screenshot shows a nearly empty scene with no visible dragon and only one distant ring dominating the frame, so the core flight gameplay reads as broken/off-camera at this moment. Great chrome, weak visual payoff.
▶ LIVEGame
Game
Renders cleanly with polished neon 3D visuals, glowing gradient title, grid arena, and functional falling-orb catch mechanic. Strong presentation but gameplay is shallow/generic — no lives, misses, difficulty ramp, or lose condition, so it lands short of the task's best entries.
▶ LIVENeonblaster
Game
Strong neon HUD/title and clean starfield render with a polished glow aesthetic, and the source covers waves, bosses, power-ups, screen-shake, and synth SFX. However the screenshot shows a sparse, static-looking scene with a tiny player ship and no enemies/action visible, and the 'synth music' is really just one-shot tones rather than a soundtrack, keeping it below the field's best.
▶ LIVENeoncity
Game
Strong neon typography and clean UI, and it does render a driveable 3D city, but the buildings read as flat matte blocks (emissive materials don't self-glow without bloom) and the road looks like a plain blue plane, giving a flat, unpolished cyberpunk feel well short of the field's best.
▶ LIVENeonracer
Game
Strong neon aesthetic with glowing title, cyan/pink pillars and a clear car, and the code has a working follow-cam, boost and vapor particles; but the screenshot shows almost no visible particle trail and the pillars float/scatter awkwardly, leaving it polished-looking but generic rather than a standout racer.
▶ LIVENordiccrypt
Game
Renders with a strong Cinzel title and clean HUD, but the scene is blown out in flat orange — torch point lights over-saturate everything so pillars/floor lose all stone texture and the 'crypt' reads as a bright yellow tunnel rather than a dark torch-lit ruin. Camera also spawns clipped inside a pillar; atmosphere and mood miss the brief badly.
▶ LIVEOutrun
Game
Renders cleanly with a striking neon title, glowing cyan road with dashed center line, and colorful side buildings, but the flat-shaded matte cubes read more like a generic 3D scene than a synthwave city, and there's no sun/horizon gradient or visible car — polished UI but modest atmosphere.
▶ LIVEPool
Game
Renders a clean 3D table with rack, pockets, and a decent title overlay, but the table sits small in the frame with heavy vignette wasting most of the screen, and the physics have flaws (cue-ball-only drag, no game rules/scoring). The rack looks slightly offset and the balls are plain-colored with no numbers/stripes, keeping it generic rather than a task winner.
▶ LIVERaycaster
Game
Polished gradient title/HUD and clean colorful walls, but the initial render looks flat and disorienting — walls float in a black void with no visible floor/ceiling, and the camera appears to be facing into a wall rather than presenting a readable corridor. The build uses full 3D meshes with collision (functional, walkable) but the framing/lighting fails to sell a convincing Wolfenstein-style maze on first view.
▶ LIVERpg
Game
Polished UI chrome (title, inventory grid, hint) renders nicely, but the actual game world is empty—no player sprite, enemies, or items visible, and the inventory slots are blank, suggesting the camera/scene or sprite spawn failed. As an RPG with 'sprites, combat, inventory' the core gameplay simply isn't showing.
▶ LIVEMatrixrain 🥈
Other
Strong glowing Orbitron title and a legible 3D perspective character field with good fog/depth make it polished and clearly on-brief, but the falling glyphs read as sparse scattered planes rather than dense classic Matrix columns, and much of the frame is empty black — solid and shippable but not quite the field's best.
▶ LIVEMlx Speedtest 🥈
Other
Strong atom/orbit-ring 3D visual with a polished shimmer title and clean readout typography, and it clearly renders on-brand. However the 'speedtest' is purely cosmetic random noise with no real benchmark logic, and the big readout awkwardly overlaps the core sphere, keeping it just shy of the field's best.
▶ LIVENeonsnake 🥉
Other
Strong glowing title typography and clean grid/UI, but the neon snake reads as a small dim segmented tube with weak emissive glow and minimal presence—the centerpiece lacks the vivid neon punch and visual scale the brief demands, leaving it below the field's best.
▶ LIVELanding
Page
Clean, polished dark landing with nice gradient mesh, glass cards with per-card glow accents, and a strong CTA button — but the hero H1 is clipped by the parallax offset so only 'Brand' shows (the 'Illuminate Your Brand' headline is cut off at top), plus the header/nav is scrolled out of frame, which hurts first impression. Solid execution, generic-modern copy, and the truncated headline keeps it below the top of the field.
▶ LIVEWebos
Page
Renders cleanly with polished dock, desktop icons, and a functional Terminal window with prompt; drag, close/minimize dots, Paint canvas and localStorage Notes all present per source. Weak point: the title banner is partially hidden behind the window and the empty terminal body looks sparse, making it feel slightly less finished than the top field entries.
▶ LIVEBlackhole
Sim
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.
▶ LIVEBoids
Sim
Renders cleanly with a polished neon title, grid, starfield and working 3D boid flocking, but the sparse scattered cones and one oversized foreground boid look under-clustered and slightly awkward rather than showing tight emergent flocks; solid but not a task winner.
▶ LIVECloth
Sim
Renders cleanly with polished typography/UI, but the core brief fails: the 'cloth' sits as a flat plane hovering entirely above the sphere with no actual draping, and the sphere is fully visible rather than covered — it's a sine-wave sheet, not physical cloth simulation with real constraints or drape.
▶ LIVEFluid
Sim
Renders cleanly with a polished Orbitron title and a dense particle cloud with pleasing color mix, but it reads as a scattered particle sphere rather than convincing swirling fluid motion — the flow field is subtle and the CPU per-particle loop limits it to a generic point cloud rather than the cohesive vortex the brief implies.
▶ LIVEFractal
Sim
Clean GPU-shader Mandelbrot renders crisply with smooth iteration coloring, glow, and polished title/hint overlay plus proper pan+zoom with lerp smoothing. Loses ground on being pan-only (no zoom-to-cursor), a slightly muddy blue-orange palette, and lacking Julia/coordinate readouts that top entries offered.
▶ LIVEGalaxy
Sim
Clean gradient title and multi-color particle disc render nicely, but the galaxy is pushed to the bottom of the frame with a hollow center and hard square particles (no soft round sprites), and the mouse only rotates rather than truly swirling — decent but not a task winner.
▶ LIVEOrbit
Sim
Clean render with polished title, colorful glowing bodies, starfield and working 3D orbit/spawn interactions; but the physics uses non-symplectic Euler with a hard bounds hack, no orbital trails, and the scattered layout doesn't visually read as gravitational clustering—competent but generic versus the field's best.
▶ LIVEParticleforge
Sim
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.
▶ LIVEPathtracer
Sim
Genuine Monte Carlo path tracer with cosine-weighted sampling and accumulation, and the neon title/UI is polished — but the render is dominated by a blown-out overexposed emissive sphere that washes over everything, the other spheres are barely visible, and the tone-mapping (sqrt without proper 255 scaling clamp) looks broken, making the scene read as a poor image rather than a convincing render.
▶ LIVEReactiondiff
Sim
Strong header/UI polish but the simulation fails the core brief: instead of intricate Turing patterns it shows a uniform pink blob with no structure — the seed/color mapping saturated B everywhere and the classic Gray-Scott spots/stripes never emerge. The stretched aspect ratio (canvas is full-screen but sim is square) also distorts the field.
▶ LIVESolar
Sim
Renders cleanly with polished typography, orbit rings, starfield and interactive controls, but the camera framing is broken — the Sun sits off-center, planets aren't clearly the largest bodies, no Saturn rings, and the scale/order reads as a generic sphere scatter rather than an accurate, legible solar system.
▶ LIVEWormhole
Sim
Renders cleanly with polished pastel rings, nice typography, and clear HUD, but the spiral path produces a chaotic tangle of loops rather than a coherent tunnel-through-the-center flythrough — it reads as scattered torus rings, not a wormhole you fly down.
▶ LIVEAurora
Visual
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.
▶ LIVEFireworks
Visual
Polished neon Orbitron title and clean glassmorphic hint render well, and the Three.js particle/explosion system with raycasted click-launch and auto-spawn is solid; but the screenshot shows an empty sky with no visible fireworks yet, so it reads generic and unproven versus the field's best animated captures.
▶ LIVELavalamp
Visual
Renders cleanly with warm glowing spheres and elegant typography, but the discrete shiny orbs read as floating balls rather than a true lava lamp — no metaball merging, no vertical rise/fall, and no glass vessel breaks the brief's core morph concept.
▶ LIVEMatrix 🥈
Visual
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.
▶ LIVEPlasma
Visual
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.
▶ LIVESynthwave
Visual
Strong typography and neon grid perspective land the synthwave vibe, but the sun renders as a flat pale-white disc (not a warm gradient sunset) and the emissive cyan 'lines' read as stray beams shooting through the sun rather than a polished retro loop, keeping it below the field's best.
▶ LIVETerrain
Visual
Renders and is clearly 3D terrain with orbit controls and decorative boxes, but the camera is zoomed too close and the washed-out pale-blue coloring reads as an overexposed ice blob rather than a legible landscape; per-vertex re-computation every frame is also costly. Strong plumbing, weak final visual composition.
▶ LIVEVoxel
Visual
Renders a clean 3D voxel terrain with decent lighting, shadows, and polished UI overlay, but the random per-cube color scattering reads as noisy rather than coherent Minecraft-style biomes (grass/dirt/stone layers), and the drag-rotate control is disconnected from the actual auto-orbiting camera—solid but generic.
▶ LIVEWaves
Visual
Renders cleanly with a polished cyan gradient ocean, glowing WAVES title, and floating droplet spheres, but the wave surface reads as a smooth blob rather than crisp animated swells and the low camera angle flattens the effect. Competent and on-brief but lacks the specular sparkle and detailed wave definition of the top entries.
every demo, in a grid · click any one to play
Compare Inkling against every other model
Every head-to-head featuring Inkling. Verdicts shown for scored pairs.
Inkling
vs
Fusion
Fusion leads 45–2
Inkling
vs
Hermes MoA
Hermes MoA leads 44–2
Inkling
vs
GPT-5.6 Sol
GPT-5.6 Sol leads 48–2
Inkling
vs
Claude Fable 5
Claude Fable 5 leads 45–1
Inkling
vs
Grok
Grok leads 39–4
Inkling
vs
MiniMax M3
MiniMax M3 leads 44–3
Inkling
vs
Fugu Ultra
Fugu Ultra leads 37–5
Inkling
vs
GLM-5.2
GLM-5.2 leads 43–4
Inkling
vs
Fugu Mini
Fugu Mini leads 30–6
Inkling
vs
Opus 4.8
Opus 4.8 leads 41–6
Inkling
vs
Kimi K2.7
Kimi K2.7 leads 19–6
Inkling
vs
Qwable 5 27B Coder
Qwable 5 27B Coder leads 27–12
Inkling
vs
Claude Sonnet 5
Claude Sonnet 5 leads 32–13
Inkling
vs
Qwen 3.7
Qwen 3.7 leads 34–13
Inkling
vs
Agents-A1
Inkling leads 25–14
Inkling
vs
Gemma 4 12B · MLX
Inkling leads 34–6
Inkling
vs
Laguna XS 2.1
Inkling leads 35–5
Inkling
vs
Qwythos 9B
Inkling leads 38–4
Inkling
vs
LongCat-2.0
LongCat-2.0 leads 4–0
Inkling
vs
Hy3
Hy3 leads 5–1
Inkling
vs
Gemma-4 12B Coder
Inkling leads 5–1
Inkling
vs
Kimi K2.7 · Fast
47 shared tasks · unscored
Inkling
vs
Kimi K2.7 · No-Think
47 shared tasks · unscored
Inkling
vs
Kimi K2.7 · Quality
47 shared tasks · unscored
Inkling
vs
Ornith 1.0
42 shared tasks · unscored
Inkling
vs
Claude Mythos 5
Reference-only
Inkling
vs
Kilo Code
Reference-only
See all 66 comparisons across every model →
Quick pill index
Direct comparisons against every other scored model on the bench:
Inkling vs Fusion Inkling vs Hermes MoA Inkling vs GPT-5.6 Sol Inkling vs Claude Fable 5 Inkling vs Grok Inkling vs MiniMax M3 Inkling vs Fugu Ultra Inkling vs GLM-5.2 Inkling vs Fugu Mini Inkling vs Opus 4.8 Inkling vs Kimi K2.7 Inkling vs Qwable 5 27B Coder Inkling vs Claude Sonnet 5 Inkling vs Qwen 3.7 Inkling vs Agents-A1 Inkling vs Gemma 4 12B · MLX Inkling vs Laguna XS 2.1 Inkling vs Qwythos 9B Inkling vs LongCat-2.0 Inkling vs Hy3 Inkling vs Gemma-4 12B CoderRead more on agentos.guide: /ink-machine
Inkling — frequently asked
What is Inkling?
Inkling is Thinking Machines's AI model — A 975B open-weights frontier model — yours to own and run. It has a 1M tokens context window and was released 2026-07.
How good is Inkling at coding and one-shot builds?
On the GoldieBench one-shot build benchmark it averages 6.07/10 across 50 scored tasks, with 0 gold, 3 silver and 1 bronze medals.
How much does Inkling cost?
$0.33 / M. 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/
Where can I see Inkling demos?
Every one-shot build is live and playable on this page and on the GoldieBench compare matrix — same prompt as every other model, no retries.
The same stack Julian uses
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.
4,000+founders
258documented wins
38countries
$59/momonthly