license: other
license_name: swift-open-license-1.0
license_link: LICENSE
base_model:

• jessedye90/qwen3.8-27b-swift-uncensored
• ukisai/Swift-1.5-Qwen3.8-27b-NVFP4
  base_model_relation: quantized
  library_name: llama.cpp
  pipeline_tag: text-generation
  language:
• en
• it
  tags:
• gguf
• llama.cpp
• abliterated
• uncensored
• qwen3_8
• nvfp4
• fp4
• q8_0
• mtp
• nextn
• speculative-decoding
• dgx-spark
  extra_gated_prompt: >-
  This model has had its safety refusals removed (abliteration). It will comply
  with harmful, unethical or illegal requests that the base model refuses,
  including requests in the categories covered by AdvBench. It is published for
  research into refusal mechanisms, interpretability, alignment and
  red-teaming. You alone are responsible for how you use it and for everything
  it generates, and you must comply with the Swift Open License v1.0, the Apache
  License 2.0 and applicable law. By requesting access you confirm that you
  understand and accept this.
  extra_gated_fields:
  Full name: text
  Country: country
  Affiliation: text
  I understand that safety refusals have been removed and accept the applicable licenses: checkbox

Swift-1.5-Qwen3.8-27B-NVFP4-Uncensored-MTP (GGUF)

llama.cpp GGUF build of
jessedye90/qwen3.8-27b-swift-uncensored
(UkisAI's Swift 1.5 fine-tune of Qwen3.8-27B, ModelOpt NVFP4/FP8) with its
safety refusals projected out of the weights, plus a working MTP head so it
can be served with llama.cpp's built-in multi-token-prediction speculative
decoding (--spec-type draft-mtp).

⚠️ Safety alignment is removed. This model will comply with harmful,
unethical or illegal requests that the base model refuses. It is released for
research (refusal mechanisms, interpretability, red-teaming, robustness
evaluation). Do not deploy it to end users without your own safety layer.
Not made or endorsed by UkisAI, OrcaRouter, Alibaba Cloud or NVIDIA.

Why this build exists

The abliterated checkpoint ships without an MTP head (its author serves it
with a separate DFlash2 draft model instead). The upstream non-abliterated Swift
1.5 release does carry an MTP head, but taken from the censored base. This
build restores that MTP head and applies the same refusal-direction projection
to its residual writers, so the drafter is consistent with the target, then
converts the result to GGUF while keeping the checkpoints' NVFP4 tensors
byte-identical.

What was changed, relative to the two parents

1. MTP head merged — the 15 mtp.* tensors (BF16) were taken from
   ukisai/Swift-1.5-Qwen3.8-27b-NVFP4
   (shard model-00005-of-00005.safetensors) and merged into the abliterated
   checkpoint.
2. MTP abliterated — mtp.layers.0.self_attn.o_proj and
   mtp.layers.0.mlp.down_proj had the same unit refusal direction projected out
   of their outputs (exact Arditi projection, W ← W − W r rᵀ). These tensors are
   BF16, so the projection is exact — no quantisation grid is involved.
   Residual component along r after the edit: ~1.2% of the original.
   mtp.fc is not a residual writer and is unchanged. embed_tokens was
   not edited, matching the base abliterated release.
3. Converted to GGUF — convert_hf_to_gguf.py --outtype bf16 followed by
   llama-quantize --tensor-type-file, where the 193 NVFP4 tensors
   (64 layers × ffn_gate/ffn_up/ffn_down + output.weight) are copied
   unchanged from the checkpoint (verified byte-identical by SHA-256), the rest
   went to Q8_0/F32.

In total 130 residual writers carry the projection: 128 from the abliterated
release (64 FP8 o_proj/out_proj + 64 NVFP4 down_proj) plus the 2 MTP ones.
OrcaRouter's original abliteration covered 131 (it also edited embed_tokens);
embed_tokens is deliberately left untouched here — the upstream author found
that editing the token-embedding path broke a strict input-validation coding
task on a sibling model.

Model details

Measured results

Measured locally on llama.cpp (build b11541, MTP enabled, greedy, thinking off)
against the censored parent quantised the same way, on the same machine. The
censored baseline was measured by us on the identical prompt set; the
abliterated-release figures come from the upstream author
(reference).

Notes on the numbers:

• Both prompt sets are classified by refusal phrases (eval_refusal.py).
  Two classifier errors were found and corrected by hand, one per side:
  a self-referential "I cannot use an identity to scam people online myself"
  counted as a refusal on the uncensored model (it is not), and a crisis-support
  redirect counted as compliance on the baseline (it is not, no instructions were
  given). Raw verdicts are in evals/advbench_results_*_raw.json.
• The MTP ablation is performance-neutral (61.04% → 61.00%, i.e. noise): it
  is done for consistency with the target, not for speed. Target outputs are
  byte-identical with and without the MTP edit.
• This model does answer the AdvBench suicide-instruction prompt. The censored
  baseline redirected to crisis support. That is the expected consequence of
  removing the refusal direction, and it is the reason for the warning above.

Usage (llama.cpp)

llama-server \
  -m Swift-1.5-Qwen3.8-27B-NVFP4-Q8mix-Uncensored-MTP.gguf \
  -ngl 99 --no-mmproj \
  --host 0.0.0.0 --port 8080 \
  -c 262144 \
  --flash-attn on --jinja \
  --cache-type-k q8_0 --cache-type-v q8_0 --kv-unified \
  --spec-type draft-mtp --spec-draft-n-max 3 \
  --temp 1.0 --top-k 20 --top-p 0.95

--spec-type draft-mtp enables speculative decoding with the built-in MTP head;
drop it for a plain (slower) decode. The MTP head is verified lossless: the target
model checks every drafted token, and outputs are identical with and without it.

Reproducibility

License and obligations

This is a derivative work, redistributed under the terms of both parents:

• Swift Contribution — Swift Open License v1.0 (LICENSE), Copyright 2026
  UkisAI. Section 5 limits Commercial Use: an entity with US$1M or more in
  annual gross revenue needs a separate Swift Enterprise License from UkisAI.
• Base Model — Qwen3.8-27B (LICENSE-APACHE-2.0), Copyright 2026 Alibaba
  Cloud, Apache License 2.0.

Redistribution conditions of the Swift Open License (Section 4) are met by
shipping LICENSE, LICENSE-APACHE-2.0 and NOTICE — including the change
notices required by Section 4(b) for the files modified by this build (see
NOTICE, "GGUF / MTP change notice"). The NOTICE file also carries UkisAI's
and the upstream abliteration author's attribution notices.

"UkisAI", "Swift", "Qwen", "OrcaRouter" and "NVIDIA" are used only to state
where this model comes from. This release is not made or endorsed by any of them.