GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
ctxnn1
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
ctxnn1
Model Tree
Qwen/Qwen3-8B
Input Modalities
Output Modalities
Supported Functionality
On 10 held-out subjects (never trained on), 4 seeds each, rated blind by me (love / okay / no):
| Model | Valid code | Loved | Strict mean (0–2) |
|---|---|---|---|
| v2 best (cats from raw shapes) | 38 / 40 | 1 | 0.30 |
| Qwen3-8B + kit, untrained | 14 / 40* | 13 | 0.68 |
| v3 SFT | 36 / 40 | 28 | 1.57 |
| v3 SFT + RL (this adapter) | 39 / 40 | 33 | 1.75 |
* Most failures came from one kit effect (weather('petals')) that crashed the headless renderer;
it has been fixed in the catkit.js shipped here.
An automatic check (Qwen2.5-VL asked "does this painting show <subject>?") says the painting clearly shows the requested subject in 28 / 40 held-out paintings (v2: 5 / 40).
0.10 + 0.40 · sigmoid(HPSv3 μ / 4) + 0.50 · P(Qwen2.5-VL says it shows the subject); else 0.python
from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B", revision="b968826d9c46dd6066d109eabc6255188de91218", torch_dtype="bfloat16", device_map="auto")model = PeftModel.from_pretrained(base, "ctxnn1/reipaint-qwen3-8b-v3-catkit")tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B", revision="b968826d9c46dd6066d109eabc6255188de91218") prompt = PROMPT # see "Prompt" below; replace the subjecttext = tok.apply_chat_template([{"role": "user", "content": prompt}], tokenize=False, add_generation_prompt=True, enable_thinking=False)out = model.generate(**tok(text, return_tensors="pt").to(model.device), do_sample=True, temperature=1.0, top_p=1.0, top_k=0, max_new_tokens=1024)print(tok.decode(out[0], skip_special_tokens=True))To see the painting, load p5.js 2.3.2, p5.brush 2.2.1 and catkit.js in a page, then the
generated code (it defines setup() and draw() on a 600 × 600 WEBGL canvas):
html
Generated code should only ever be run in a sandbox. The training renderer first checks it against an allowlist (only the kit's functions, numbers and strings; no arrays, objects, network or DOM).
The exact training prompt, here for the held-out subject "a cat reading a tiny book":
text
Write JavaScript that paints a cute, simple watercolour illustration in the style of a Pinterest cat drawing: soft round shapes, gentle pastel colours, a plain cream background. The subject: a cat reading a tiny book. Compose the picture with the cat kit below: choose the pose, coat, eyes, props, weather and layout that best show this exact subject, and make it cute. Code rules: only the kit functions, setup/draw as in the template, let/const numbers and strings, for loops, if, arithmetic, sin(), cos(), random() and Math functions. No arrays, objects, text or other functions. Kit functions (call them inside draw, in painting order: background first, cat, then what goes on top): paper(tone) paper background, call first. tone optional, e.g. '#fff6ea' backdrop(kind, colour) kind: 'circle' 'arch' 'hill' 'window' 'sunset' 'night' ground(colour) soft floor shadow cat(x, y, size, pose, coat, eyes, eyeColour) x, y = where the cat sits (bottom centre); size about 0.5 to 1.1 pose: 'sit' 'kitten' 'loaf' 'sleep' 'side' 'lie' 'wave' 'stretch' 'back' 'peek' coat: 'orange' 'ginger' 'cream' 'grey' 'black' 'white' 'calico' 'tabby' 'siamese' 'tuxedo' 'silhouette' or a '#hex' colour eyes: 'open' 'wide' 'happy' 'sleepy' 'wink' 'up' 'yawn'; eyeColour optional '#hex' wear(kind, colour) on the last cat: 'scarf' 'bow' 'partyhat' 'knithat' 'crown' 'flower' hold(kind, colour) in the last cat's paws: any item kind catIn(kind, x, y, size, coat, eyes, colour) a cat peeking out of: 'teacup' 'box' 'pot' 'basket' 'knithat' 'mug' item(kind, x, y, size, colour) 'yarn' 'book' 'books' 'butterfly' 'flower' 'fish' 'fishbowl' 'cupcake' 'heart' 'moon' 'sun' 'cloud' 'cushion' 'blanket' 'umbrella' 'watermelon' 'pumpkin' 'leaf' 'star' 'mushroom' 'mug' 'cocoa' 'plant' weather(kind, colour) 'snow' 'rain' 'stars' 'hearts' 'petals' 'leaves' 'sparkles'The canvas is 600 x 600; a cat of size 1 sitting at (300, 500) fills about half of it. Template:```javascriptfunction setup() { createCanvas(600, 600, WEBGL); angleMode(DEGREES); brush.scaleBrushes(3); noLoop(); }function draw() { translate(-width / 2, -height / 2); paper(); // your picture here: kit calls with numbers and strings}``` Return exactly one ```javascript code block and nothing else.from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B", revision="b968826d9c46dd6066d109eabc6255188de91218", torch_dtype="bfloat16", device_map="auto")model = PeftModel.from_pretrained(base, "ctxnn1/reipaint-qwen3-8b-v3-catkit")tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B", revision="b968826d9c46dd6066d109eabc6255188de91218") prompt = PROMPT # see "Prompt" below; replace the subjecttext = tok.apply_chat_template([{"role": "user", "content": prompt}], tokenize=False, add_generation_prompt=True, enable_thinking=False)out = model.generate(**tok(text, return_tensors="pt").to(model.device), do_sample=True, temperature=1.0, top_p=1.0, top_k=0, max_new_tokens=1024)print(tok.decode(out[0], skip_special_tokens=True))Write JavaScript that paints a cute, simple watercolour illustration in the style of a Pinterest cat drawing: soft round shapes, gentle pastel colours, a plain cream background. The subject: a cat reading a tiny book. Compose the picture with the cat kit below: choose the pose, coat, eyes, props, weather and layout that best show this exact subject, and make it cute. Code rules: only the kit functions, setup/draw as in the template, let/const numbers and strings, for loops, if, arithmetic, sin(), cos(), random() and Math functions. No arrays, objects, text or other functions.Kit functions (call them inside draw, in painting order: background first, cat, then what goes on top): paper(tone) paper background, call first. tone optional, e.g. '#fff6ea' backdrop(kind, colour) kind: 'circle' 'arch' 'hill' 'window' 'sunset' 'night' ground(colour) soft floor shadow cat(x, y, size, pose, coat, eyes, eyeColour) x, y = where the cat sits (bottom centre); size about 0.5 to 1.1 pose: 'sit' 'kitten' 'loaf' 'sleep' 'side' 'lie' 'wave' 'stretch' 'back' 'peek' coat: 'orange' 'ginger' 'cream' 'grey' 'black' 'white' 'calico' 'tabby' 'siamese' 'tuxedo' 'silhouette' or a '#hex' colour eyes: 'open' 'wide' 'happy' 'sleepy' 'wink' 'up' 'yawn'; eyeColour optional '#hex' wear(kind, colour) on the last cat: 'scarf' 'bow' 'partyhat' 'knithat' 'crown' 'flower' hold(kind, colour) in the last cat's paws: any item kind catIn(kind, x, y, size, coat, eyes, colour) a cat peeking out of: 'teacup' 'box' 'pot' 'basket' 'knithat' 'mug' item(kind, x, y, size, colour) 'yarn' 'book' 'books' 'butterfly' 'flower' 'fish' 'fishbowl' 'cupcake' 'heart' 'moon' 'sun' 'cloud' 'cushion' 'blanket' 'umbrella' 'watermelon' 'pumpkin' 'leaf' 'star' 'mushroom' 'mug' 'cocoa' 'plant' weather(kind, colour) 'snow' 'rain' 'stars' 'hearts' 'petals' 'leaves' 'sparkles'The canvas is 600 x 600; a cat of size 1 sitting at (300, 500) fills about half of it.Template:```javascriptfunction setup() { createCanvas(600, 600, WEBGL); angleMode(DEGREES); brush.scaleBrushes(3); noLoop(); }function draw() { translate(-width / 2, -height / 2); paper(); // your picture here: kit calls with numbers and strings}```Return exactly one ```javascript code block and nothing else.