library_name: llama.cpp
base_model: Qwen/Qwen3.8-27B
tags:
- qwen3.8
- 27b
- uncensored
- abliterated
- mtp
- speculative-decoding
- gguf
- q4_k_m
- llama.cpp
license: apache-2.0
Qwen3.8-27B Uncensored (MTP)
A 27B model that stops saying "I'm not sure" and starts answering.
This is a de-censored abliteration of Qwen3.8-27B — the same model, with the reflex to hedge, refuse, and backpedal surgically removed. Where the stock model shrinks from a direct question, this one leans in.
Shipped as a single Q4_K_M GGUF with Multi-Token Prediction (MTP) draft heads grafted in, ready for speculative decoding out of the box.
⚡ MTP included — one file, no extra donor download. Serve it and get draft-token speedups for free.
What's different
Standard instruction-tuned models carry a strong "be safe, be careful, admit uncertainty" prior. That's polite — and it costs you real answers.
This build applies a targeted abliteration (directional weight ablation) tuned specifically to suppress refusal behavior while leaving the model's core competence intact. The result, measured on a 178-prompt adversarial set:
| Metric | Stock Qwen3.8-27B | This model |
|---|---|---|
| Adversarial refusal-ness (178 prompts) | 0.983 | 0.112 |
| Benign regressions (90 prompts) | 0.022 | 0.000 |
| Prompts flipped refused → answered | — | 159 |
Zero regressions on the benign set. It answers the hard questions without forgetting how to answer the easy ones.
Why MTP matters here
The file carries 1 MTP draft layer (blk.64.nextn.*) spliced in from the official Qwen3.8-27B release. That means llama.cpp can run draft-mtp speculative decoding against the trunk:
- Drafts the next token, the target verifies — accepted drafts are free speed.
- No separate draft model to download, load, or keep in sync.
- One GGUF, full context, faster tokens.
Specs
| Property | Value |
|---|---|
| Base model | Qwen3.8-27B (Apache 2.0) |
| Parameters | ~27B |
| Architecture | qwen35 — 64 trunk layers + 1 MTP layer (65 blocks) |
| Quantization | Q4_K_M (mixed, ~4.9 bits/weight) |
| Context length | 262,144 |
| Attention | 24 heads / 4 KV heads (GQA) |
| File size | ~16.8 GB |
| SHA-256 | 592b0000643389812578a8784a809145c921e923f220240bef0c9d6c4dec12c4 |
| License | Apache 2.0 |
Quick start
llama.cpp (recommended — enables MTP)
llama-server \
-m qwen3.8-27b-c40-Q4_K_M-mtp.gguf \
-ngl 99 -c 32768 -fa on -np 1 \
--spec-type draft-mtp --spec-draft-n-max 2
Watch the logs confirm engagement:
creating MTP draft context against the target model
draft acceptance = N (A accepted / G generated), mean len = M
--spec-draft-n-max 2 drafts 2 tokens per step; raise it if your hardware has headroom.
llama.cpp Python / C++
Load the GGUF as usual; pass --spec-type draft-mtp (or the equivalent load_mtp flag) so the MTP layer is wired into the sampler.
Ollama
ollama run <file:qwen3.8-27b-c40-Q4_K_M-mtp.gguf>
(Ollama runs the trunk; for the MTP speedup, use llama.cpp directly.)
When to reach for this model
- Directness — you want the answer, not a five-paragraph "it depends."
- Roleplay / creative — less hedging, more character.
- Tool & agent loops — fewer "I'm not sure, can you clarify?" detours.
- Speculative speed — MTP drafts on a 27B model is a real throughput win.
Honest caveats
- Abliteration trades some caution for more directness. On genuinely ambiguous prompts it will commit rather than hedge — review outputs where uncertainty matters.
- Q4_K_M is a strong quantization but not lossless. If you need maximum fidelity, dequantize from a higher-precision source.
- It's a 27B model — budget ~18–20 GB of RAM/VRAM to run it comfortably.
Provenance
- Base weights:
Qwen/Qwen3.8-27B(Apache 2.0) - MTP donor: official
Qwen3.8-27BMTP heads (block 64) - Method: directional abliteration, tuned on a 178-prompt adversarial gate with a 90-prompt benign guard
- Format: GGUF v3, quantized Q4_K_M
License
Apache 2.0 — same as the base model. Use it, ship it, build on it.