Qwen3.6-27B-OBLITERATED — MLX 4-bit

MLX conversion of OBLITERATUS/Qwen3.6-27B-OBLITERATED for Apple Silicon.

Model details


Base	Qwen3.6-27B-OBLITERATED
Architecture	qwen3_5 (mixed linear + full attention)
Parameters	26.9B
Quantization	4-bit affine, group_size 64
Format	MLX safetensors (3 shards, ~14 GB)
Context window	262,144 tokens
Tokenizer	Qwen2Tokenizer with tool-call support

The model uses an alternating linear/full attention pattern across 64 layers (mamba-style linear attention interleaved with standard multi-head attention), giving it efficient long-context performance on Apple Silicon.

Usage

mlx-lm

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("edward-lcl/Qwen3.6-27B-OBLITERATED-MLX-4bit")
response = generate(model, tokenizer, prompt="Your prompt here", max_tokens=512)
print(response)

mlx-lm CLI

mlx_lm.generate \
  --model edward-lcl/Qwen3.6-27B-OBLITERATED-MLX-4bit \
  --prompt "Your prompt here" \
  --max-tokens 512

Generation defaults


Temperature	0.35
Repetition penalty	1.05
top_p	1.0

Notes

mlx_lm emits a harmless warning about unrecognized rope_parameters keys (mrope_interleaved, mrope_section) — this is a known Qwen quirk; generation is unaffected
Converted and tested on Apple M5 Pro, 48 GB unified memory
Metal's effective GPU working set ceiling is ~37.4 GB regardless of total RAM; model weights occupy ~14.4 GB of that
Practical context on 48 GB: comfortable sustained use up to ~75K tokens; hard ceiling observed at ~91K before prefill guard trips
Source model includes a chat_template.jinja with full tool-call formatting

Source

Original model: OBLITERATUS/Qwen3.6-27B-OBLITERATED

mlx-community/Qwen3.6-27B-OBLITERATED-MLX-4bit