Modelpublic

h1st0ry3D/Godoter-27B-MLX-4bit

sourceHugging Faceotherupdated 3mo agoView on Hugging Face

1likes190downloads

Model Card

Godoter-27B-MLX-4bit

MLX 4-bit quantized version of Ruler97/Godoter-27B — a Qwen3.5-27B finetune specialized for Godot 4 GDScript game development.

Converted with Unsloth 2026.6.6 and quantized to 4-bit (group_size=64, affine mode) for compatibility with oMLX and Apple MLX on Apple Silicon.

Property	Value
Base model	unsloth/Qwen3.6-27B
Finetune	Ruler97/Godoter-27B — Godot GDScript expert
Architecture	Qwen3.5 (dense 27B, 64 layers, 24 attention heads, 4 KV heads)
Context length	262,144 tokens (theoretical), 32,768 recommended on 32 GB
Quantization	4-bit, group_size=64, affine mode
Weight size	~14 GB (3 safetensors shards)
Format	MLX safetensors (`mlx` library)
MTP	1 MTP prediction head (included in weights)

File	Size	Description
`config.json`	4.2 KB	Architecture + quantization config
`model-00001-of-00003.safetensors`	5.0 GB	Weight shard 1
`model-00002-of-00003.safetensors`	5.0 GB	Weight shard 2
`model-00003-of-00003.safetensors`	4.1 GB	Weight shard 3
`model.safetensors.index.json`	185 KB	Weight map index
`tokenizer.json`	19 MB	Tokenizer
`tokenizer_config.json`	1.2 KB	Tokenizer config
`chat_template.jinja`	7.9 KB	Qwen3.5 chat template
`generation_config.json`	213 B	Generation defaults

—This is a pure text model (the vision tower from Qwen3.5-VL is excluded from the quant — only the text decoder was converted)
—The model has MTP (Multi-Token Prediction) heads but can be run without them
—For 64k context, use the GGUF version instead (llama.cpp handles KV cache more efficiently for dense models)
—Converted with Unsloth 2026.6.6 (unsloth_fixed_mtp: true)