LongShu
A large language model fine-tuned specifically for game development, acting as a technical director and lead architect that breaks down game system requirements and produces an executable development plan.
Built on a Qwen3.5-122B-A10B mixture-of-experts architecture with domain-specific fine-tuning and engine-aware quantization, it runs at about 3.5 tokens per second on consumer hardware, such as a 64GB Apple Silicon Mac or two RTX 3090/4090 GPUs, while keeping 98.5% of its base reasoning ability, according to the project's own benchmark.
Its 60-layer network alternates linear and full attention in a 3:1 ratio, natively supports a 262,144-token context window, and sparsely activates only 10 of its 105 experts per token, with the 20 most important experts kept at higher precision and the rest compressed aggressively; it is designed to output stable JSON for coordinating downstream agents rather than to chat.
It was trained on more than 50 real game projects, over 2 billion tokens of Unreal C++, Unity C# and Godot/Lua source, and more than 300,000 pages of game design documents, and is run locally through llama.cpp with a published 4-bit MLX quantization.