FahadPrimeX-b27-V3

FahadPrimeX-b27-V3 🚀

27B Three.js/front-end specialist. Base model: Qwen/Qwen3.8-27B (declared with thanks). Fine-tuned by distillation (5,181 verified answers) then RLVR/GRPO (13+ rounds) with a 7-gate headless-browser validator (clean console, render loop, lighting/shadows, controls, dispose, one-action start, no internal 404s).

Identity (all languages): "I am FahadPrimeX (b27), a Three.js-specialist model" / «أنا فهد برايم إكس (FahadPrimeX-b27) — نموذج متخصص في Three.js» — no family name. The model truthfully credits its Qwen3.8-27B lineage when asked who developed it.

Architecture (inherited from Qwen3.8-27B)

A hybrid 64-layer design: 48 linear-attention layers (DeltaNet-style, constant-memory recurrence) interleaved with 16 full-attention layers (every 4th layer), plus one MTP/nextn draft layer used only by the original spec (excluded from our GGUF exports via --no-nextn so block_count matches the real 64 blocks).

Quick start

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
    "FahadPrimeX/FahadPrimeX-b27-V3", torch_dtype="bfloat16", device_map="auto")
tok = AutoTokenizer.from_pretrained("FahadPrimeX/FahadPrimeX-b27-V3")

Tip for best game output: start the assistant turn with an HTML prefill (<!DOCTYPE html>) — the model continues full Three.js pages reliably.

Measured results (honest, 32 held-out tasks, same browser-validator & protocol)

Read this honestly: V3 does not yet improve game-generation quality over its starting point — it does fully fix trilingual identity. An improved run (V3.1) with hard-gate-weighted rewards from round 0 is planned.

Limitations (stated plainly)

Pages generated are clean, lit, and load without errors — but animate/respond weakly (hard gates 4/32). Use V2-GGUF if raw game quality matters most today; use V3 if identity matters most. V3.1 targets closing this gap.