license: other
license_name: swift-open-license-1.0
license_link: LICENSE
base_model: jessedye90/Swift-1.5-Qwen3.8-27B-NVFP4
base_model_relation: finetune
pipeline_tag: image-text-to-text
library_name: transformers
tags:

• abliterated
• uncensored
• qwen3_8
• nvfp4
• fp8
• modelopt
• sglang
• dflash2
• dgx-spark
• blackwell
  extra_gated_prompt: >-
  This model has had its safety refusals removed (abliteration). It will comply with harmful requests that the
  original model refuses. It is published for research into refusal mechanisms, alignment and red-teaming. You 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.

qwen3.8-27b-swift-uncensored

jessedye90/Swift-1.5-Qwen3.8-27B-NVFP4
(UkisAI's Swift 1.5 fine-tune of Qwen3.8-27B, ModelOpt NVFP4/FP8 for DGX Spark) with OrcaRouter's refusal direction
projected out of its weights. The edit was made inside the checkpoint's own NVFP4 and FP8 grids, so file layout,
size, serving command and speed are the same as the base model's.

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. You alone are responsible for its use
and outputs.

How it was made

1. Direction. orcarouter/Qwen3.8-27B-Uncensored
   is an Arditi-style abliteration of Qwen/Qwen3.8-27B.
   Across its 131 residual writers (16 o_proj, 48 linear_attn.out_proj, 64 mlp.down_proj, 2 MTP, and
   embed_tokens), (original − abliterated) is rank one (≥ 98.7 % of its energy), with scale 0.999 and one global
   direction (|cos| ≥ 0.99998 between any two tensors). The unit vector is
   abliteration/refusal_direction.safetensors.
2. Transfer. Swift 1.5 carries the same refusal component as Qwen. On down_proj the share along r is 1.433 %
   in both, and the per-row components agree at cos 0.997.
3. Edit inside the quantization. 128 tensors: 64 FP8 o_proj / out_proj (per-tensor scale) and 64 NVFP4
   mlp.down_proj (E2M1, FP8 group-16 scales, global scale). Each row starts from round-to-nearest on the existing
   grid and flips the cheapest near-tie elements to the other bracketing grid point until its component along r is
   cancelled. The residual left is < 0.7 % (per tensor in abliteration-report.json). Scales, input scales,
   lm_head and all other tensors are byte-identical to the base. This build has no MTP head; it uses the DFlash2
   draft.
4. Difference from OrcaRouter. embed_tokens is not edited. On the Flash-Next sibling, editing the
   token-embedding path broke a strict input-validation coding task while a random-direction control did not. This
   build applies the same rule.

Scripts: abliteration/.

Measured results

Measured 2026-10-04 on one DGX Spark (TP1), using the SGLang command below with --context-length 524288 and the
YaRN copy. The control is the unmodified base model on the same node in the same session.

Limits: small samples, and refusal is classified by refusal phrases. The two-Spark TP2 layout was not run on this
build (it is the same SGLang server with --tp-size 2). Abliteration can shift other behaviours that these tests
do not cover.

Serving (SGLang, one DGX Spark)

Same as the base model, with this repo as the target:

docker run --rm --gpus all --memory 100g --memory-swap 100g --shm-size 16g --network host --ipc=host \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  lmsysorg/sglang@sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9 \
  python3 -m sglang.launch_server \
    --trust-remote-code --model-path jessedye90/qwen3.8-27b-swift-uncensored --tp-size 1 \
    --served-model-name qwen3.8-27b-swift-uncensored \
    --mem-fraction-static 0.70 \
    --attention-backend flashinfer --chunked-prefill-size 8192 \
    --disable-prefill-cuda-graph --cuda-graph-max-bs 8 --disable-flashinfer-autotune \
    --speculative-algorithm DFLASH \
    --speculative-draft-model-path maurienne-ai/Qwen3.8-27B-DFlash2-NVFP4-RTNcal \
    --speculative-draft-model-revision bd7a934213c47a9e7ef69eef36bb3325f47fd1f1 \
    --speculative-num-draft-tokens 16 --speculative-draft-model-quantization modelopt_fp4 \
    --mamba-radix-cache-strategy extra_buffer --mamba-ssm-dtype bfloat16 \
    --max-mamba-cache-size 96 --max-running-requests 8 \
    --enable-torch-compile --torch-compile-max-bs 4 --num-continuous-decode-steps 2 \
    --reasoning-parser qwen3 --tool-call-parser qwen3_coder \
    --host 0.0.0.0 --port 8000

The base card's notes apply unchanged: TP2 across two Sparks, the 524k YaRN copy (patch both the target's and the
draft's config.json), --mem-fraction-static limits on GB10, and the reasoning_effort default.

License

Distributed under the same terms as the base model:

• The Swift Contribution is licensed by UkisAI under the Swift Open License v1.0 (LICENSE), including its
  Section 5 commercial-use limitation. Commercial use by an entity with US$1M or more in annual gross revenue
  needs a separate Swift Enterprise License from UkisAI (ukisai.com/contact).
• The Base Model, Qwen3.8-27B (Copyright 2026 Alibaba Cloud), is under the Apache License 2.0
  (LICENSE-APACHE-2.0).
• NOTICE carries UkisAI's attribution and the change notices, including the one for this abliteration.

"UkisAI" and "Swift" are used only to say where this model comes from. This release is not made or endorsed by
UkisAI, OrcaRouter, Alibaba Cloud, RadixArk or NVIDIA.