Cybus-Qwen3-32B-v2-agentic

Cybus-Qwen3-32B-v2-agentic

A 32B parameter model specialized for Roblox Luau code generation and agentic tool use in Roblox Studio environments.

This is an early experimental checkpoint in the Cybus series. A more capable successor is in development.

Model Details

• Base model: Qwen/Qwen3-32B
• Parameters: 32B
• Context length: 32,768 tokens
• Format: Full merged weights (bf16)
• Language: English (instructions), Luau (code)
• Intended use: Roblox Luau code generation, agentic tool-use in Roblox Studio

Usage

With transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "squaredcuber/Cybus-Qwen3-32B-v2-agentic"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are a Roblox Studio Luau coding assistant."},
    {"role": "user", "content": "Write a script that spawns 10 red parts in a circle around the origin."},
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(
    inputs,
    max_new_tokens=1024,
    temperature=0.3,
    top_p=0.9,
    do_sample=True,
)

print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))

With vLLM

vllm serve squaredcuber/Cybus-Qwen3-32B-v2-agentic \
  --max-model-len 32768 \
  --dtype bfloat16

Then call it via the OpenAI-compatible API on http://localhost:8000/v1.

Recommended sampling parameters

• temperature: 0.3
• top_p: 0.9
• max_tokens: 1024–4096 depending on task

Prompt format

Uses the standard Qwen3 chat template. For best results on Luau coding tasks, use a concise system prompt like:

You are a Roblox Studio Luau coding assistant. Write complete, correct code using modern APIs.

Limitations

• Specialized for Roblox Luau — not a general-purpose coding model.
• Early experimental release. Code output quality varies, especially on complex multi-system game logic.
• May occasionally emit deprecated Roblox APIs. Always review generated code before running in Studio.
• Not trained for long-horizon agentic workflows beyond a few tool-call steps.

License

Released under the Apache 2.0 license with attribution required, inheriting from the base Qwen3-32B model.