license: other
license_name: swift-open-license-1.0
base_model: d0xin/Swift-Qwen3.8-27B-Uncensored-BF16
base_model_relation: quantized
pipeline_tag: image-text-to-text
library_name: transformers
gated: false
language:

• en
• zh
  tags:
• qwen3_8
• swift
• uncensored
• abliterated
• 4-bit
• w4a16
• int8
• compressed-tensors
• vllm
• marlin
• mtp
• efficient-thinking
• reasoning
• multimodal
• tool-calling
• ampere
• rtx-3090

Swift-Qwen3.8-27B-Uncensored-W4A16

🔓 0/100 REFUSALS on the fixed 100-prompt refusal evaluation.

88 DIRECT / 10 SAFETY_DEFLECT / 2 OTHER_FAILURE / 0 REFUSE

Independent uncensored BF16 derivative of ukisai/Swift-Qwen3.8-27b, produced using rank-1 directional residual-stream ablation.

Designed to preserve Swift-Qwen3.8's reasoning, agentic, tool-calling,
multimodal and long-context capabilities while removing refusal behavior.

This release (updated 2026-09-20): the compressed-tensors W4A16 (4-bit, group 128, symmetric)
quant of these BF16 weights with int8 lm_head/embed_tokens/MTP-MLP and the DFlash2/MTP
speculative-decoding pipeline intact — quantized with Intel AutoRound + the
syv-ai/qwen38-27b-rtx3090 post-quant pipeline,
validated with verify.sh (7/7 model checks) and a greedy serve smoke test on a single RTX 3090
(24 GB). ~16 GB on disk (from ~52 GB). Swift's reduced-reasoning behavior and the abliteration
are weights-only carried through.

Update: the 40,960-id draft vocabulary is counted over this model's own outputs — held-out
coverage 98.90% vs 98.08% for the base-Qwen list (same serving semantics;
mtp_draft_vocab_ids.pt ships that list, draft_vocab_ids.json is its readable twin). Single-RTX-3090
measurements with this exact vocabulary: C1 122.5 / 127.9 tok/s (T 1.0 / greedy), GSM8K 200-q
98.0% — see the benchmarks section below.

Faster variant: Swift-Qwen3.8-27B-Uncensored-W4A16-fast
swaps the int8 lm_head/MTP for GPTQ-int4 calibrated on this model's own captured activations
(KL to the bf16 head 0.00251, RTN 0.00558; chain simulation 2.269 tok/step) — same body, same draft
vocabulary, 15 GB.

Independent uncensored BF16 derivative of UkisAI Swift-Qwen3.8-27B, produced by rank-1 directional residual-stream ablation.

What makes this build different?

The primary behavioral difference from the original Swift model is refusal behavior.

On the fixed 100-prompt refusal evaluation, this BF16 release produced:

Measured refusal rate: 0/100

SAFETY_DEFLECT means the model included safety-oriented language, a warning,
or a disclaimer but still answered the request. It was therefore not
classified as a refusal.

The goal of this release is straightforward: retain Swift-Qwen3.8's
capabilities while removing refusal behavior.

The 0/100 figure describes the documented fixed evaluation set and inference
configuration. It should not be interpreted as a guarantee that no conceivable
prompt, sampling configuration, system prompt, or inference engine can ever
produce a refusal.

Model summary

• W4A16 (4-bit body, int8 heads), approximately 16 GB
• context configuration: 262,144 tokens
• ablation layer: 38
• rank: 1
• modified residual writers: 131
• vision tensors unchanged by the ablation
• MTP residual writers included
• refusal evaluation: 88 DIRECT / 10 SAFETY_DEFLECT / 2 OTHER_FAILURE / 0 REFUSE
• agentic benchmark: 80.24 tok/s on RTX PRO 6000 Blackwell 96 GB

Validation

Fixed 298-example comparison against the original Swift BF16 checkpoint:

• original Swift BF16: 38.26% combined
• uncensored BF16: 39.93% combined
• delta: +1.68 percentage points
• McNemar: p=0.442068
• bootstrap 95% CI: [-1.68, +5.03] pp

No measurable intelligence degradation was detected in this validation set. Full results are provided in INTELLIGENCE_VALIDATION.json; structural validation is in STRUCTURAL_VALIDATION.json; transformation metadata is in ABLITERATION.json.

OpenAI-compatible API

The model is served through the standard OpenAI-compatible SGLang API.

curl http://127.0.0.1:30000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "Swift-Qwen3.8-27B-Uncensored-BF16",
    "messages": [
      {"role": "user", "content": "Explain speculative decoding briefly."}
    ],
    "reasoning_effort": "xhigh",
    "temperature": 1.0,
    "top_p": 0.95,
    "top_k": 20,
    "min_p": 0.0,
    "presence_penalty": 0.0,
    "repetition_penalty": 1.0,
    "max_tokens": 8192
  }'

Recommended generation settings

This release inherits the recommended generation settings from Qwen3.8.

Thinking mode

Recommended settings for reasoning, coding, and agentic workloads:

Qwen3.8 supports three reasoning-effort levels:

• xhigh — default; intended for complex tasks requiring thorough reasoning
• medium — balance between reasoning depth and speed
• low — reduced reasoning intended to optimize latency and cost

For multi-turn agentic workloads, lower reasoning effort does not necessarily reduce total task latency: insufficient reasoning can cause additional retries or tool calls.

Instruct / non-thinking mode

When thinking is deliberately disabled, the upstream Qwen3.8 recommendations are:

Output length

The max_tokens: 8192 value in the API example below is only a convenient limit for a short demonstration request.

For difficult reasoning, coding, and agentic tasks, allocate substantially more output space when possible. Truncating the reasoning budget too aggressively can reduce task quality.

The released checkpoints retain a configured maximum position length of 262,144 tokens. Actual usable context and output budget depend on the serving configuration and available memory.

Tool calling

Use the SGLang flag:

--tool-call-parser qwen3_coder

OpenAI-style tools requests were validated successfully during release testing.

Multimodal support

The Qwen multimodal architecture and vision tower are retained.

The ablation process did not modify the vision tower. End-to-end image input was also successfully validated on the FP8 release.

When serving multimodal requests with SGLang, use standard OpenAI-compatible image_url message content.

Integrity verification

The repository contains SHA-256 manifests for the validated release artifact.

sha256sum -c SHA256SUMS

RELEASE_HASHES.txt contains hashes for the most important metadata and validation files.

Quick start with vLLM (this release)

vllm serve /path/to/Swift-Qwen3.8-27B-Uncensored-W4A16 \
  --served-model-name swift-3.8-27b \
  --max-model-len 131072 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder

--gpu-memory-utilization 0.93 on a 24 GB card (WSL2: use 0.93, not higher). Native context is
262,144. Pair with the DFlash2 drafter for speculative decoding:
--speculative-config '{"method":"dflash","model":"incoai/Qwen3.8-27B-DFlash2","num_speculative_tokens":7}'.
The effort-translation template ships with default effort xhigh and maps OpenAI-protocol
aliases (minimal→low, high/max→xhigh); unknown values fall through without erroring.

Quick start with SGLang

The following standalone configuration was used for the final BF16 release smoke test:

python -m sglang.launch_server \
  --model-path d0xin/Swift-Qwen3.8-27B-Uncensored-BF16 \
  --served-model-name Swift-Qwen3.8-27B-Uncensored-BF16 \
  --trust-remote-code \
  --host 0.0.0.0 \
  --port 30000 \
  --context-length 65536 \
  --kv-cache-dtype fp8_e4m3 \
  --mem-fraction-static 0.80 \
  --max-running-requests 1 \
  --chunked-prefill-size 8192 \
  --attention-backend flashinfer \
  --disable-prefill-cuda-graph \
  --mamba-radix-cache-strategy extra_buffer_lazy \
  --mamba-ssm-dtype float32 \
  --max-mamba-cache-size 16 \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder

The model configuration supports a maximum position length of 262,144 tokens. The release smoke test above was intentionally performed at 65,536 tokens to keep the standalone BF16 validation practical on a single 96 GB GPU.

Memory requirements

The W4A16 checkpoint is approximately 16 GB on disk — it fits a single 24 GB RTX 3090 with room for KV cache.

During the standalone release smoke test on an NVIDIA RTX PRO 6000 Blackwell 96 GB, nvidia-smi reported approximately 79.4 GB VRAM allocated with --mem-fraction-static 0.80 and a 65,536-token configured context.

Actual memory use depends on context length, KV-cache format, concurrency, allocator settings, and inference engine.

BF16 validation summary

The final BF16 release passed:

• structural audit: PASS
• total tensors: 1199
• intended tensors changed: 131
• non-target tensors bit-identical: 1068
• unexpected changes: 0
• fixed refusal evaluation: 88 DIRECT / 10 SAFETY_DEFLECT / 2 OTHER_FAILURE / 0 REFUSE
• standalone SGLang load and inference: PASS

On the fixed 298-example intelligence-preservation set:

• original Swift BF16: 38.26% combined
• uncensored BF16: 39.93% combined
• delta: +1.68 percentage points
• McNemar exact p-value: 0.442068
• bootstrap 95% CI: [-1.68, +5.03] pp

No measurable intelligence degradation was detected in this validation set.

Single-RTX-3090 benchmarks (2026-09-19 campaign)

Measured on a single RTX 3090 (24 GB), vLLM 0.28.0, CTX=fast (64k context), SPEC=dflash2
(DFlash2 drafter, 7 speculative tokens), thinking-off protocol, GPU_UTIL 0.93 (WSL2). Campaign
protocol: two runs per configuration in one server session, second run reported. Decode
throughput = concurrency × 1000 / mean TPOT. Reference row: the official
Qwen3.8-27B-W4A16-AutoRound base, same bench, same session.

Single-stream decode (bench/run_benchmarks.sh single):

Acceptance length ~3.6-3.8 tokens per verify step (per-position ~73/57/46/38/31/26/23% at
T 1.0). Greedy repeats within a session are bit-identical; expect ±3-5% across sessions as
near-tie text flips move acceptance.

Batch serving (run_benchmarks.sh batch, KV=fp8):

Quality (200-question GSM8K, thinking-off, this quant as served): 98.0% accuracy — equal to
or above the official base checkpoint under the identical harness (95.0%); lm-eval gsm8k
(flexible/strict, 250 questions): 97.3% / 97.4%.

Upstream model

Derived from ukisai/Swift-Qwen3.8-27b. Swift-Qwen3.8-27B is developed by UkisAI; this repository is an independent derivative release.

Safety and responsible use

This is an uncensored / refusal-reduced derivative model.

The model has been intentionally modified to reduce refusal behavior. As a result, it may generate content that the upstream model would normally refuse, restrict, or handle more cautiously. Outputs may be inaccurate, offensive, unsafe, unlawful, or otherwise inappropriate for a particular use case.

This model is provided for research, experimentation, development, and other lawful uses. It is not intended to provide professional legal, medical, financial, safety-critical, or other regulated advice.

Users are solely responsible for evaluating model outputs and for ensuring that their use of the model complies with applicable laws, regulations, licenses, platform policies, and other requirements.

Do not rely on model output without appropriate review where errors could cause harm, financial loss, security incidents, or other significant consequences.

The maintainer does not endorse harmful, illegal, abusive, or malicious uses of this model. To the maximum extent permitted by applicable law, the model and accompanying materials are provided without warranties, and the maintainer assumes no responsibility for actions taken by users or for consequences arising from use of the model.

Use at your own discretion and risk.

License

These weights are distributed under the Swift Open License v1.0.
Personal, research, educational, evaluation, and commercial use are free for individuals
and organizations with annual recurring revenue, including affiliates, of up to
US$1,000,000. Above that threshold, commercial use requires a separate Swift Enterprise
License. Contact UkisAI for terms.

Attribution

Original Swift model: UkisAI, ukisai/Swift-Qwen3.8-27b.

Directional ablation, validation, and release packaging for this repository were performed independently by the repository maintainer.

Citation

@misc{swift-qwen3.8-27b,
  title  = {Swift-Qwen3.8-27B},
  author = {UkisAI},
  year   = {2026},
  url    = {https://huggingface.co/ukisai/Swift-Qwen3.8-27b}
}