Qwen3.8-27B heretic-ara — oQ3e (MTP preserved)
Mixed-precision quant of trohrbaugh/Qwen3.8-27B-heretic-ara, produced with oQ (oMLX v0.6.2). 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 3-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 3-bit.
Performance (oMLX on M4 10-core)
| Context | PP tok/s | TG tok/s | Peak Mem |
|---|---|---|---|
| 1k | 63.0 | 11.1 | 15.5 GB |
| 4k | 61.8 | 12.3 | 17.1 GB |
With Lightning MTP enabled. For reference, with MTP disabled the same quant measures ~8.0 / 7.7 TG tok/s at 1k / 4k (13.4 / 14.9 GB) — on the lower-bit quants the MTP head mostly costs memory rather than adding throughput.
| Batch | TG tok/s | Speedup |
|---|---|---|
| 1× | 11.1 | 1.00× |
| 2× | 14.5 | 1.31× |
| 4× | 24.8 | 2.23× |
Full benchmark → (1k) · 4k
Why this quant
The original BF16 weights require ~55 GB. This oQ3e quant runs in ~13–15 GB on Apple Silicon while keeping the vision tower and MTP head intact — the lightest option in the oQe family.
Quick start
# oMLX
omlx serve --model underlotus/Qwen3.8-27B-heretic-ara-oQ3e-mtp
# mlx-lm
from mlx_lm import load, generate
model, tokenizer = load("underlotus/Qwen3.8-27B-heretic-ara-oQ3e-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.