Get the Agent OS + join 4,000+ founders inside the AI Profit Boardroom → Join AIPB ($59/mo)
Richard Young · DeepNeuro (abliterated build of empero-ai's Qwythos, Qwen3.5 base)

Qwythos 9B

A Claude-style creative & reasoning 9B with a full 1M-token context — the local writer & thinker.

Context1,048,576 tokens (YaRN, Qwen3.5-9B base)
PricingFree · runs locally
Tasks tested42
Avg score2.98/10 average
Medals🥇0 🥈0 🥉0
Release2026-06
Official vendor source
Qwythos 9B is built by Richard Young · DeepNeuro (abliterated build of empero-ai's Qwythos, Qwen3.5 base) — see the vendor's own product page, pricing, and docs at ollama.com/richardyoung/qwythos-9b-abliterated.
Visit ollama.com/richardyoung/qwythos-9b-abliterated →

What is Qwythos 9B?

Qwythos 9B is the Richard Young · DeepNeuro (abliterated build of empero-ai's Qwythos, Qwen3.5 base) frontier model with a 1,048,576 tokens (YaRN, Qwen3.5-9B base) context window, released 2026-06. Tagline: A Claude-style creative & reasoning 9B with a full 1M-token context — the local writer & thinker.. Official source: ollama.com/richardyoung/qwythos-9b-abliterated.

Pricing detail. Qwythos (model ID: richardyoung/qwythos-9b-abliterated) is an abliterated build of empero-ai's Qwythos-9B-Claude-Mythos — a Claude-style creative & reasoning model on a Qwen3.5-9B base, post-trained on Claude Mythos & Fable traces, with a 1M-token context. Thinking model (<think>), native function-calling. Refusals trimmed via the Heretic library (53/100, KL 0.0066). Q4_K_M is 5.6GB, runs 100% offline.

How I use it inside the Agent OS. Loaded in Ollama as the local long-context + creative option. Now fully bench-scored (42/42, avg 2.98): it handles simpler page and arcade builds, but most of its ambitious one-shot 3D/WebGL attempts errored out — an honest read on a free 9B's ceiling for hard one-shots.

What I built with Qwythos 9B

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 Qwythos 9B shipped on the bench: 42 one-shot demos across 1,048,576 tokens (YaRN, Qwen3.5-9B base) of context. Of those, 42 are scored against the field with my honest 0–10 from the source guides at agentos.guide.

Strengths

  • Full 1M-token context locally — matches frontier-cloud models on context length, at $0
  • Fast on a Mac — a steady ~52 tokens/sec on real builds (M4 Max), just 5.6GB on disk
  • Open-weights (Qwen3.5 base), runs 100% on-device via Ollama — free and private
  • Best bench builds: Nordic Crypt 6.8, Racing 6.5, Neon Blaster 6.4, Arcade 6.2, Landing 6.0

Trade-offs

  • 9B ceiling — bench-scored 2.98/10 avg (42/42), last on the one-shot creative-coding suite
  • Ambitious 3D/WebGL one-shots often ship with real bugs — bare `three` imports, const reassignment, undefined refs — so they render blank
  • Strong on simpler page/arcade/2D tasks; falls apart on complex graphics where frontier models hold up

Best for

  • Long-context local tasks (large file refactors, multi-file analysis) where 1M ctx matters
  • Cost-zero daily coding work on a consumer Mac
  • Workflows where you want an unfiltered local coder without API guardrails

Every demo by Qwythos 9B

42 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.

Qwythos 9B one-shot build of Crypt — GoldieBench AI benchmark screenshot▶ LIVE
Crypt
Game
The start screen is clean and atmospheric with a nice monospace title treatment, but the screenshot shows only the menu — no actual dungeon renders, and the source has serious bugs (Player/Enemy reference this.rooms which never exists, broken collision damaging HP on every frame), so the core 3D crawler likely doesn't work.
Qwythos 9B one-shot build of Dragonrealm — GoldieBench AI benchmark screenshot▶ LIVE
Dragonrealm
Game
Only the HUD panel renders — the 3D scene is empty black with no ground, crystals, snow, or runes visible, indicating the Three.js world failed to display (likely the camera/scene setup or CDN issue). The HUD styling is decent but the core 3D dragonrealm experience is broken.
Qwythos 9B one-shot build of Doom — GoldieBench AI benchmark screenshot▶ LIVE
Doom
Game
The canvas renders pure black with no raycast view at all — only the title and empty HUD bars show, meaning the core engine fails to draw. The map layout uses inconsistent row lengths and the canvas appears never sized, so the central deliverable is broken.
Qwythos 9B one-shot build of Racing — GoldieBench AI benchmark screenshot▶ LIVE
Racing
Game
Clean neon title menu with gradient text and a glowing start button, but the screenshot shows only the menu over a black void — no visible 3D track or car, and the source has a clear bug (CONFIG.acceleration is undefined, so acceleration logic is NaN). Polished landing screen but unproven gameplay caps it below the strong field.
Qwythos 9B one-shot build of Skyrim — GoldieBench AI benchmark screenshot▶ LIVE
Skyrim
Game
The 3D scene failed to render — only the title, controls hint, and crosshair appear over an empty background color, with no terrain, trees, or water visible. Likely a SimplexNoise import or canvas init failure broke the Three.js setup.
Qwythos 9B one-shot build of Twilightvale — GoldieBench AI benchmark screenshot▶ LIVE
Twilightvale
Game
The build is stuck on the loading spinner — the 3D world never renders, almost certainly because PointerLockControls is referenced as THREE.PointerLockControls (not imported) and game.enemy is undefined, throwing errors that prevent the loader from clearing. Decent ambition in the source, but the screenshot shows a non-functional product.
Qwythos 9B one-shot build of Voxelcraft — GoldieBench AI benchmark screenshot▶ LIVE
Voxelcraft
Game
The screenshot shows only a loading spinner on a black background — the voxel world never rendered, so the build is effectively broken on screen. The source has ambitious intent (Three.js terrain, chunks, day/night) but the loader never clears and the UI/world failed to appear.
Qwythos 9B one-shot build of Arcade — GoldieBench AI benchmark screenshot▶ LIVE
Arcade
Game
The neon title and start button render with a nice retro arcade aesthetic, but the game container is tiny and mis-sized, with the title and button overflowing the playfield and the score UI cramped/illegible — the layout looks broken rather than polished, and the convoluted touch-swipe handling is buggy.
Qwythos 9B one-shot build of Dogfight — GoldieBench AI benchmark screenshot▶ LIVE
Dogfight
Game
The output dumped raw planning text and thinking notes as page content instead of producing a working game — the screenshot shows only unstyled prose, no canvas, no Three.js scene, nothing playable.
Qwythos 9B one-shot build of Dragonflight — GoldieBench AI benchmark screenshot▶ LIVE
Dragonflight
Game
The HUD chrome (title, fury meter, fire-ready indicator, controls) renders cleanly with nice neon styling, but the 3D scene is completely empty/black — no dragon, rings, grid, or lights are visible, so the core dragonflight experience failed to render.
Qwythos 9B one-shot build of Game — GoldieBench AI benchmark screenshot▶ LIVE
Game
Game
The neon title/UI and color palette read nicely, but the physics is badly broken: particles fly off the top and pile against the ceiling, leaving the screen mostly empty with faint smeared trails — the gravity sign and force math are inverted, so it isn't a fun or working game.
Qwythos 9B one-shot build of Neonblaster — GoldieBench AI benchmark screenshot▶ LIVE
Neonblaster
Game
Clean neon start screen with a nicely glowing cyan button and HUD scaffolding, but the screenshot only shows the menu — no gameplay visible — and the source has questionable particle/bullet logic (trail check kills bullets after 1 unit) that suggests the actual game may not play smoothly.
Qwythos 9B one-shot build of Neoncity — GoldieBench AI benchmark screenshot▶ LIVE
Neoncity
Game
The 3D scene fails to render — only the HUD overlays (title + controls hint) show on a black void, with no grid, buildings, or car visible. The code has broken camera/animation math (uninitialized gridHelper.material.map causing a likely runtime error) and the city never renders, leaving a completely empty result.
Qwythos 9B one-shot build of Neonracer — GoldieBench AI benchmark screenshot▶ LIVE
Neonracer
Game
Only the title/subtitle UI renders over a black void — the 3D scene is entirely absent, likely due to runtime errors (e.g. invalid mat.setColor/rotateZ chained on materials and old.mesh.dispose on a Group). Nothing of the actual racer is visible or playable.
Qwythos 9B one-shot build of Nordiccrypt — GoldieBench AI benchmark screenshot▶ LIVE
Nordiccrypt
Game
Clean, atmospheric start screen with a nicely styled title card, slanted button and HUD controls bar, but the screenshot only shows the menu — the actual 3D crypt isn't visible, and the source has questionable code (CanvasTexture with no image) that suggests the 3D scene may be fragile or unrendered.
Qwythos 9B one-shot build of Outrun — GoldieBench AI benchmark screenshot▶ LIVE
Outrun
Game
The synthwave grid floor and title styling show intent, but the render is dominated by a flat magenta sky covering the top half with no horizon sun, road, or car visible — the title overflows off-screen and the start screen looks broken rather than polished. The pseudo-3D road code is convoluted and the screenshot suggests the actual driving scene never renders correctly.
Qwythos 9B one-shot build of Pool — GoldieBench AI benchmark screenshot▶ LIVE
Pool
Game
The 3D scene fails to render entirely — only the title and hint text appear over a blank dark background, with no table, balls, or pool game visible. The source has fatal issues like calling sphereGeo()/ballMat() before they're defined and truncated physics, so the canvas stays empty.
Qwythos 9B one-shot build of Raycaster — GoldieBench AI benchmark screenshot▶ LIVE
Raycaster
Game
Renders a fully black screen — the build is broken: width/height are never set so the canvas/camera are zero-sized, and update() references undefined variables (dx, dz, rayDist) that throw. The raycasting logic is also incoherent, mixing a Three.js scene with confused pixel manipulation.
Qwythos 9B one-shot build of Rpg — GoldieBench AI benchmark screenshot▶ LIVE
Rpg
Game
The start screen renders cleanly with a decent neon title and clear instructions, but the source contains fatal errors — THREE.Mesh().position.set() returns undefined (so nothing is added to scene), and THREE.ShadowMap is not a real constructor — meaning the actual RPG game almost certainly never initializes past this menu.
Qwythos 9B one-shot build of Landing — GoldieBench AI benchmark screenshot▶ LIVE
Landing
Page
Clean dark hero with nice gradient headline, glassy feature cards, and a solid CTA, but the centerpiece Three.js scene fails to render (the JS references an undefined `scene` and never creates one), leaving a mostly empty black canvas that undercuts the 'immersive 3D' promise.
Qwythos 9B one-shot build of Webos — GoldieBench AI benchmark screenshot▶ LIVE
Webos
Page
Desktop icons and taskbar render cleanly, but the Three.js background failed to appear (likely the truncated/broken resize handler or CDN issue) leaving a flat black void, and no windows are shown — the centerpiece 3D aesthetic and app functionality aren't visible in the actual render.
Qwythos 9B one-shot build of Blackhole — GoldieBench AI benchmark screenshot▶ LIVE
Blackhole
Sim
The screenshot shows only UI text on a black void with no black hole, accretion disk, or stars rendered — the core 3D scene failed to appear, likely due to the overly complex/broken CPU raymarching texture logic overdrawing everything.
Qwythos 9B one-shot build of Boids — GoldieBench AI benchmark screenshot▶ LIVE
Boids
Sim
Only the title/instructions UI renders — no boids, no grid, no 3D content visible, indicating the Three.js scene failed to display (likely the boids spawn out of camera view or render order issues). The code structure is reasonable but the actual rendered result is essentially empty.
Qwythos 9B one-shot build of Cloth — GoldieBench AI benchmark screenshot▶ LIVE
Cloth
Sim
Only the UI text label renders — the canvas is empty black with no visible cloth or terrain, indicating the WebGL scene failed to display (likely the physics spring force explodes/NaNs the geometry on frame one). The code also has a buggy spring sign and uses a nonexistent geometry.setXYZ method, so the simulation is effectively broken.
Qwythos 9B one-shot build of Fluid — GoldieBench AI benchmark screenshot▶ LIVE
Fluid
Sim
Only the UI text overlay renders on a black canvas — the fluid simulation is completely absent. The shaders are fundamentally broken (vUv used before assignment in vertex shader, wrong attribute access via getAttribLocation, uniformMatrix4fv called on a float uniform), so nothing draws.
Qwythos 9B one-shot build of Fractal — GoldieBench AI benchmark screenshot▶ LIVE
Fractal
Sim
The fractal never renders — only the title/instructions overlay shows on a black screen, because the color packing uses bit-shifts (col.r<<16) on already-decomposed channel values, producing garbage/zero pixels. Despite the ambitious Three.js + canvas setup, the core Mandelbrot output is broken on screen.
Qwythos 9B one-shot build of Galaxy — GoldieBench AI benchmark screenshot▶ LIVE
Galaxy
Sim
Only the UI overlay renders — the galaxy itself is invisible, leaving an empty black screen. The shader declares uniforms (size, color) as uniforms but they're set as per-vertex attributes/never bound, so the vertex shader fails or produces zero-size points; the brief's core visual never appears.
Qwythos 9B one-shot build of Orbit — GoldieBench AI benchmark screenshot▶ LIVE
Orbit
Sim
The build leaked its entire chain-of-thought as visible text above a tiny black canvas, and the rendered simulation is empty/broken — the edge-fade fillRect overwrites everything each frame and the canvas isn't sized to fullscreen, so no orbiting bodies are visible.
Qwythos 9B one-shot build of Particleforge — GoldieBench AI benchmark screenshot▶ LIVE
Particleforge
Sim
Only the title overlay renders against an empty green void — no particles are visible. The BufferGeometry is constructed with an invalid constructor signature (passing attributes object directly instead of setAttribute), so the point cloud and trails never appear.
Qwythos 9B one-shot build of Pathtracer — GoldieBench AI benchmark screenshot▶ LIVE
Pathtracer
Sim
The UI panel renders but the canvas is entirely black — no path-traced scene appears, and the source is broken (invalid GLSL calling JS Noise.simplex inside the shader, a top-level `return` in module scope, and nonsensical raytracing math). Nothing of the actual brief renders.
Qwythos 9B one-shot build of Reactiondiff — GoldieBench AI benchmark screenshot▶ LIVE
Reactiondiff
Sim
The UI chrome (title, sliders, reset button) renders cleanly, but the actual reaction-diffusion canvas is essentially blank dark — the fragment shader uses a broken non-iterative noise hack rather than a real RD simulation, so no Turing patterns appear. The core brief is unfulfilled despite the polished controls.
Qwythos 9B one-shot build of Solar — GoldieBench AI benchmark screenshot▶ LIVE
Solar
Sim
The screenshot shows only a black screen with the UI overlay — no planets, sun, or stars rendered, indicating a runtime failure (likely the truncated resize handler and a JS error in the animate loop using undefined userData.size). The source has reasonable structure but the actual build does not render the solar system.
Qwythos 9B one-shot build of Wormhole — GoldieBench AI benchmark screenshot▶ LIVE
Wormhole
Sim
The screenshot shows only the title overlay on a black void with no wormhole rendered — the shader uses an invalid texture2D(gl_FragCoord,...) call and a vertexShader referencing undeclared vUv, which throws a compile error and breaks the render loop, plus a typo (uniforms.mousePos.uniforms) that would crash. Effectively non-functional.
Qwythos 9B one-shot build of Aurora — GoldieBench AI benchmark screenshot▶ LIVE
Aurora
Visual
The screenshot is fully black — the build renders nothing, due to the vertex shader referencing vPos before assignment and uSpeed without declaring it as a uniform, causing a shader compile failure. Broken output despite a reasonable code intent.
Qwythos 9B one-shot build of Fireworks — GoldieBench AI benchmark screenshot▶ LIVE
Fireworks
Visual
The screenshot shows essentially an empty black screen with only a faint vignette and a couple of stray dots — no fireworks render on load, and the code is buggy (const size reassignment throws, broken star/twinkle logic, no actual explosion bursts). The title overlay is clean but the core effect fails to deliver the brief.
Qwythos 9B one-shot build of Lavalamp — GoldieBench AI benchmark screenshot▶ LIVE
Lavalamp
Visual
Renders only an empty dark canvas with the caption — no lamp or blobs are visible because the spawn logic is broken (blobs start dead and spawn() immediately re-sets dead=true, so no meshes are ever created and update() crashes on null). Effectively non-functional.
Qwythos 9B one-shot build of Matrix — GoldieBench AI benchmark screenshot▶ LIVE
Matrix
Visual
The signature matrix rain effect is entirely absent — the canvas renders black with only the UI text visible, since the Points material/texture and tiny FONT_SIZE-based geometry fail to produce any visible falling glyphs. Only the HUD chrome shipped; the core brief is broken.
Qwythos 9B one-shot build of Plasma — GoldieBench AI benchmark screenshot▶ LIVE
Plasma
Visual
The plasma shader failed to render — the canvas is entirely black with only the title and palette buttons visible, so the core brief (a hypnotic plasma effect) is missing. Likely a shader compile error (the GLSL redefines built-in smoothstep, which is invalid and breaks compilation).
Qwythos 9B one-shot build of Synthwave — GoldieBench AI benchmark screenshot▶ LIVE
Synthwave
Visual
The 3D scene fails to render — only the dark background and overlay text appear, no grid or shapes. The Three.js import via unpkg likely fails (incorrect CDN module path/crossorigin), and createFloatingShape passes a malformed object instead of geometry+material to Mesh, so nothing displays.
Qwythos 9B one-shot build of Terrain — GoldieBench AI benchmark screenshot▶ LIVE
Terrain
Visual
The vertex shader references undefined variables `x` and `y` (it never declares them from `position`), so the shader fails to compile and nothing renders — the screenshot is just the empty dark background with the UI label. Broken build.
Qwythos 9B one-shot build of Voxel — GoldieBench AI benchmark screenshot▶ LIVE
Voxel
Visual
The screenshot shows only the dark background and UI text with no voxel terrain rendered at all — the scene is empty, indicating a broken or non-functional build despite the elaborate source. The terrain generation/perlin noise logic appears buggy (custom noise, malformed permutation table) and clearly failed to produce any visible geometry.
Qwythos 9B one-shot build of Waves — GoldieBench AI benchmark screenshot▶ LIVE
Waves
Visual
Only the title overlay renders; the water is invisible because the custom vertex/fragment shaders are assigned to a MeshStandardMaterial as plain properties (uniforms never declared/compiled), so the wave surface fails to display. The UI typography is clean but the core 'waves' brief is broken.
every demo, in a grid · click any one to play

Compare Qwythos 9B against every other model

Every head-to-head featuring Qwythos 9B. Verdicts shown for scored pairs.

Qwythos 9B vs Fusion
Fusion leads 42–0
Qwythos 9B vs Claude Opus 5
Claude Opus 5 leads 42–0
Qwythos 9B vs Hermes MoA
Hermes MoA leads 42–0
Qwythos 9B vs GPT-5.6 Sol
GPT-5.6 Sol leads 42–0
Qwythos 9B vs Claude Fable 5
Claude Fable 5 leads 42–0
Qwythos 9B vs Qwen 3.8
Qwen 3.8 leads 36–1
Qwythos 9B vs Grok
Grok leads 38–0
Qwythos 9B vs MiniMax M3
MiniMax M3 leads 42–0
Qwythos 9B vs Fugu Ultra
Fugu Ultra leads 42–0
Qwythos 9B vs Kimi K3
Kimi K3 leads 42–0
Qwythos 9B vs GLM-5.2
GLM-5.2 leads 42–0
Qwythos 9B vs Fugu Mini
Fugu Mini leads 36–0
Qwythos 9B vs Muse Spark 1.2
Muse Spark 1.2 leads 40–2
Qwythos 9B vs Opus 4.8
Opus 4.8 leads 41–1
Qwythos 9B vs Kimi K2.7
Kimi K2.7 leads 20–0
Qwythos 9B vs Qwable 5 27B Coder
Qwable 5 27B Coder leads 39–2
Qwythos 9B vs Gemini 3.6 Flash
Gemini 3.6 Flash leads 40–2
Qwythos 9B vs Claude Sonnet 5
Claude Sonnet 5 leads 39–0
Qwythos 9B vs Qwen 3.7
Qwen 3.7 leads 42–0
Qwythos 9B vs Fugu Ultra 1.1
Fugu Ultra 1.1 leads 16–3
Qwythos 9B vs Inkling
Inkling leads 38–4
Qwythos 9B vs Grok 4.6
Grok 4.6 leads 12–4
Qwythos 9B vs Agents-A1
Agents-A1 leads 34–5
Qwythos 9B vs Gemma 4 12B · MLX
Gemma 4 12B · MLX leads 28–12
Qwythos 9B vs Laguna XS 2.1
Laguna XS 2.1 leads 29–11
Qwythos 9B vs LongCat-2.0
LongCat-2.0 leads 4–0
Qwythos 9B vs Hy3
Hy3 leads 2–0
Qwythos 9B vs Gemma-4 12B Coder
Qwythos 9B leads 3–2
Qwythos 9B vs DeepSeek V4 Flash
42 shared tasks · unscored
Qwythos 9B vs DeepSeek V4 Pro
42 shared tasks · unscored
Qwythos 9B vs Kimi K2.7 · Fast
42 shared tasks · unscored
Qwythos 9B vs Kimi K2.7 · No-Think
42 shared tasks · unscored
Qwythos 9B vs Kimi K2.7 · Quality
42 shared tasks · unscored
Qwythos 9B vs Ornith 1.0
42 shared tasks · unscored
Qwythos 9B vs Claude Mythos 5
Reference-only
Qwythos 9B 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:

Qwythos 9B vs Fusion Qwythos 9B vs Claude Opus 5 Qwythos 9B vs Hermes MoA Qwythos 9B vs GPT-5.6 Sol Qwythos 9B vs Claude Fable 5 Qwythos 9B vs Qwen 3.8 Qwythos 9B vs Grok Qwythos 9B vs MiniMax M3 Qwythos 9B vs Fugu Ultra Qwythos 9B vs Kimi K3 Qwythos 9B vs GLM-5.2 Qwythos 9B vs Fugu Mini Qwythos 9B vs Muse Spark 1.2 Qwythos 9B vs Opus 4.8 Qwythos 9B vs Kimi K2.7 Qwythos 9B vs Qwable 5 27B Coder Qwythos 9B vs Gemini 3.6 Flash Qwythos 9B vs Claude Sonnet 5 Qwythos 9B vs Qwen 3.7 Qwythos 9B vs Fugu Ultra 1.1 Qwythos 9B vs Inkling Qwythos 9B vs Grok 4.6 Qwythos 9B vs Agents-A1 Qwythos 9B vs Gemma 4 12B · MLX Qwythos 9B vs Laguna XS 2.1 Qwythos 9B vs LongCat-2.0 Qwythos 9B vs Hy3 Qwythos 9B vs Gemma-4 12B Coder

Read more on agentos.guide:

Qwythos 9B — frequently asked

What is Qwythos 9B?

Qwythos 9B is Richard Young · DeepNeuro (abliterated build of empero-ai's Qwythos, Qwen3.5 base)'s AI model — A Claude-style creative & reasoning 9B with a full 1M-token context — the local writer & thinker. It has a 1M tokens context window and was released 2026-06.

How good is Qwythos 9B at coding and one-shot builds?

On the GoldieBench one-shot build benchmark it averages 2.98/10 across 42 scored tasks, with 0 gold, 0 silver and 0 bronze medals.

How much does Qwythos 9B cost?

Free · runs locally. Qwythos (model ID: richardyoung/qwythos-9b-abliterated) is an abliterated build of empero-ai's Qwythos-9B-Claude-Mythos — a Claude-style creative & reasoning model on a Qwen3.5-9B base, post-trained on Claude Mythos & Fa

Where can I see Qwythos 9B 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