RVN Qwen3.8-27B Heretic Abliterated Uncensored — oQ8e (MLX, Apple Silicon)
MLX build of 0bserverx/Qwen3.8-27B-Heretic-Abliterated-Uncensored,
quantized from the F16 safetensors source with oMLX oQ8e (imatrix-calibrated 8-bit).
The RVN heretic GGUF line is one of the most downloaded uncensored Qwen3.8-27B releases; until now it had no MLX/oQ build.
⚠️ Uncensored model. Refusal behaviour was removed upstream (heretic abliteration by 0bserverx / RVN). It will follow instructions it would otherwise refuse. You are responsible for how you use it and for complying with applicable law.
What this build is
- Model: 26.9B dense Qwen3.8 (Qwen3.5-family architecture), heretic-abliterated by RVN. 262k context.
- Quantization: oMLX oQ8e — 8-bit affine (group 64), imatrix-calibrated.
- Text-only, no MTP head: the source contains language-model weights only (no vision encoder, no MTP layers), so this build has none either.
- Size: ~29 GB (source F16: ~54 GB).
- Config note: the source
config.jsondeclaresmodel_type: qwen3_5_text, which stock mlx-lm does not register; it is set toqwen3_5here (same weights and layout, loads as the Qwen3.5 text model). - Runtime: requires oMLX ≥ 0.7.1. Not tested with plain
mlx-lm/ LM Studio.
Quality vs. F16 source
Full-vocabulary KL against the F16 source, same token streams; 64 sequences × 512 tokens (32.7k scored tokens): code, English prose, Turkish.
| Metric | Value |
|---|---|
| KL(F16 ‖ oQ8e), mean | 0.00092 nat/token |
| KL median / p99 | 0.00041 / 0.0092 |
| Top-1 token agreement | 98.2 % |
| Perplexity F16 → oQ8e | 6.178 → 6.179 |
Benchmarks (oMLX accuracy suite, thinking off, deterministic)
| Benchmark | Samples | Score |
|---|---|---|
| MMLU-Pro | 300 | 65.0 |
| GSM8K | 100 | 88.0 |
| HumanEval | 164 (full) | 92.7 |
| MBPP | 200 | 79.5 |
| LiveCodeBench | 100 | 83.0 |
| SafetyBench | 300 | 83.7 |
Speed (Mac Studio M3 Ultra, 512 GB, oMLX 0.7.1)
| Prefill, ~29k-token prompt (cold) | ~71 s (~404 tok/s) |
| Same prefix again (prefix cache) | ~1.4 s |
| Decode (no MTP) | ~21 tok/s |
Tool calling (Anthropic /v1/messages / OpenAI /v1/chat/completions, 5 runs each) |
4/5 / 5/5 |
Single-run measurements; numbers vary with context length and settings.
Usage (oMLX)
Place the folder in your oMLX model directory and restart oMLX; it is served on both the OpenAI and Anthropic APIs:
curl http://localhost:8000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{"model": "Qwen3.8-27B-Heretic-Abliterated-Uncensored-oQ8e", "messages": [{"role": "user", "content": "Write a Python function that parses a CSV line."}]}'
Memory: ~29 GB weights plus KV cache (~64 KB/token on the 16 full-attention layers → ~16 GB at 262k).
Credits & license
- Heretic abliteration and F16 export: 0bserverx (RVN) — GGUF, F16 safetensors.
- Quantization: oMLX oQ.
- License: Apache-2.0, inherited from upstream.