underlotus/Qwen3.8-27B-heretic-ara-oQ4e-mtp

Qwen3.8-27B heretic-ara — oQ4e (MTP preserved)

Mixed-precision quant of trohrbaugh/Qwen3.8-27B-heretic-ara, produced with oQ (oMLX v0.5.8.dev1). Vision tower intact — image + video input still works. MTP head preserved. Standard MLX safetensors — compatible with oMLX, mlx-lm, LM Studio, and any MLX-capable app.

What is oQ?

Unlike uniform 4-bit quantization, oQ is a data-driven mixed-precision quantizer that calibrates per-layer sensitivity and allocates bits where they matter most. Critical layers (embeddings, LM head, the most sensitive transformer layers) are automatically promoted to higher precision, while less sensitive layers stay at 4-bit. Typical result: ~4.6 bits-per-weight.

Benchmarked on Qwen3.5-35B-A3B (oMLX project):

Benchmark	mlx-lm 4-bit	oQ4
MMLU (300)	79.7%	83.3%
TruthfulQA (300)	87.7%	88.0%
HumanEval (full)	87.2%	85.4%
MBPP (300)	71.7%	74.3%

Performance (oMLX on M4 10-core)

Context	PP tok/s	TG tok/s	Peak Mem
1k	62.4	12.7	16.7 GB
4k	60.2	12.1	18.2 GB

Batch	TG tok/s	Speedup
1×	12.7	1.00×
2×	12.0	0.94×
4×	19.3	1.52×

Full benchmark →

Why this quant

The original BF16 weights require ~55 GB. This oQ4e quant runs in ~16–18 GB on Apple Silicon while keeping the vision tower and MTP head intact.

Quick start

# oMLX
omlx serve --model underlotus/Qwen3.8-27B-heretic-ara-oQ4e-mtp

# mlx-lm
from mlx_lm import load, generate

model, tokenizer = load("underlotus/Qwen3.8-27B-heretic-ara-oQ4e-mtp")
response = generate(model, tokenizer, prompt="Hello!", max_tokens=256)
print(response)

Original model

• Base: Qwen/Qwen3.8-27B
• Decensored: Heretic v1.2.0 (custom fork) with Arbitrary-Rank Ablation (ARA) — 0/100 refusals, KL divergence 0.0535
• Vision: Full multimodal — image + video input supported
• MTP: Multi-token prediction head preserved

License

Apache 2.0, inherited from base model.