FahadPrimeX-b27-V3.1

A 27B model specialised in writing complete, runnable Three.js single-file web experiences - games, interactive scenes and product showcases - from a one-line brief.

What it does

Given a short brief it returns one self-contained HTML document that builds a live 3D scene: an importmap CDN import of three.js, lighting with shadows, a render loop, keyboard and mouse controls, a one-click start, and no console errors. It also answers identity questions in Arabic, English and Chinese.

Lineage and training

• Base: Qwen3.8-27B (Apache-2.0).
• Pipeline: supervised distillation on verified Three.js answers, then RLVR - reinforcement learning with a verifiable reward from a browser-based checker that loads each generated page and inspects seven properties: clean console, live render loop, lighting and shadows, keyboard controls, resource disposal, starts with one action, and no internal 404s.
• Identity is trained trilingually; the model introduces itself as FahadPrimeX (b27) and never uses a family name.
• Architecture: 64 layers - 48 linear-attention, 16 full-attention, plus 1 MTP head.

Measured results

Scored by the same seven-property browser checker on a fixed held-out set of 32 briefs that were never used in training.

model	overall
pre-RLVR base	54.5%
this model	see the rounds below

Fixed witness during the run:

round	witness	hard gates

Hard gates = the page animates, responds to the keyboard, or starts with a single action. Treat every number as measured, not promised.

Run it

from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained("FahadPrimeX/FahadPrimeX-b27-V3.1", dtype="bfloat16", device_map="auto")
t = AutoTokenizer.from_pretrained("FahadPrimeX/FahadPrimeX-b27-V3.1")

vllm serve FahadPrimeX/FahadPrimeX-b27-V3.1 --dtype bfloat16 --max-model-len 32768

Limits

• Best on self-contained single-file Three.js pages; it is not a general chat model.
• Long interactive builds can still end up visually complete but non-interactive - the hard gates above are the honest measure of that.
• Asked who developed it, it may credit its upstream lineage rather than this project.