Qwenity3.6-27B-msv2

Qwenity 3.6 27B

Qwenity is a specialized version of Qwen3.6-27B, fine-tuned to act as an expert assistant for Unity game development. This repository (msv2) contains the full, merged 16-bit weights — ready to load directly, with no base model or adapter required.

Model summary

Qwenity 3.6 27B is a LoRA (16-bit) fine-tune of Qwen/Qwen3.6-27B, merged back into full bfloat16 weights. It specializes the base model for the Unity engine and C# scripting while retaining the base model's general and multimodal abilities.

• Developed by: wrayy
• Base model: Qwen/Qwen3.6-27B
• Model type: Vision-language (multimodal) causal LM — qwen3_5 architecture
• Fine-tuning: LoRA at 16-bit via Unsloth + TRL (SFT), merged to full bf16
• Size / precision: ~27B parameters, bfloat16, 15 safetensors shards (~55.6 GB)
• Language: English
• License: Apache-2.0 (inherited from the base model)

Intended use

Qwenity is purpose-built to help developers build with Unity. It targets:

• C# gameplay scripting and the Unity scripting API
• The Unity Editor, GameObjects / components, and project workflow
• Rendering, physics, animation, and UI
• Unity-oriented problem solving, debugging, and tool use

It behaves as a chat / instruction-following assistant. Outside the Unity, C#, and game-development domain it falls back to its base model's general behavior and is not specifically optimized for those tasks.

How to use

Architecture note. This model reports model_type: qwen3_5 (the Qwen3.6 family). At upload time this type is not yet in released transformers (≤ 4.57.6) and the repo ships no remote code, so a plain AutoModel.from_pretrained(...) raises "Transformers does not recognize this architecture." Use one of the options below.

Option A — Unsloth (the framework this model was built and served with)

from unsloth import FastVisionModel   # FastModel also handles this VLM

model, processor = FastVisionModel.from_pretrained(
    "wrayy/Qwenity3.6-27B-msv2",
    load_in_4bit = False,   # full bf16 merged weights (set True for 4-bit)
)
FastVisionModel.for_inference(model)

See the Unsloth docs for chat-template and inference examples.

Option B — Transformers (once it supports qwen3_5)

pip install "git+https://github.com/huggingface/transformers.git"

from transformers import AutoProcessor, AutoModelForImageTextToText

model_id = "wrayy/Qwenity3.6-27B-msv2"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(model_id, torch_dtype="auto", device_map="auto")

messages = [{"role": "user", "content": [{"type": "text",
            "text": "In Unity C#, how do I make a GameObject follow the mouse cursor in world space?"}]}]
inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt",
).to(model.device)
out = model.generate(**inputs, max_new_tokens=512)
print(processor.batch_decode(out[:, inputs["input_ids"].shape[-1]:], skip_special_tokens=True)[0])

Training

• Base model: Qwen/Qwen3.6-27B
• Method: LoRA fine-tuning at 16-bit, merged to full bfloat16 weights (checkpoint 1200); trained with Unsloth + TRL supervised fine-tuning (SFT).
• Training data: wrayy/unity.masterset.sbv2 — a privately compiled and synthesized instruction dataset for Unity development (English; conversational instruction / Q&A spanning C# scripting, the Editor, rendering, physics, UI, and tool use). The dataset is private / research-only, and its contents are not publicly disclosed.

Limitations

• Domain-focused: optimized for Unity / C# / game development; it may be less reliable on unrelated tasks.
• Verify generated code: like all LLMs it can produce incorrect, outdated, or insecure code, or reference APIs that don't exist — review and test before use.
• Loading: the qwen3_5 architecture requires Unsloth or a source build of transformers (see How to use).

License

Released under Apache-2.0, inherited from the base model Qwen/Qwen3.6-27B. The training dataset is private and research-only.