kaushikvira/Qwen3.8-27B-swift15-uncensored-ajgazin-nvfp4full-dflash2-NInfer-v3

Swift-1.5 Qwen3.8-27B Uncensored (ajgazin-direction) nvfp4full + DFlash2 for NInfer — v3 container

An all-NVFP4 abliterated artifact of ukisai/Swift-1.5-Qwen3.8-27b for the NInfer engine (v3 container), with z-lab's DFlash2 speculative drafter and indexed proposal head embedded. Uncensored sibling of Qwen3.8-27B-swift15-nvfp4full-dflash2-NInfer-v3.

Credits & provenance (full chain)

All credit for the underlying model belongs to the original authors. This artifact is a format conversion + quantization + refusal-direction ablation — no fine-tuning of our own.

Role	Model	Author	SHA / commit
Source weights (BF16)	ukisai/Swift-1.5-Qwen3.8-27b	ukisai (UkisAI) — RL+OPD post-training of Qwen3.8-27B	18-shard BF16 safetensors, published 2026-09-24; shards sha256-verified against HF LFS on download
Upstream base	Qwen/Qwen3.8-27B	Qwen team (Alibaba Cloud), Apache-2.0	—
Abliteration	ajgazin/Swift-1.5-Qwen3.8-27B-Uncensored-MTP — single refusal direction recovered from orcarouter/Qwen3.8-27B-Uncensored	ajgazin / orcarouter	recovered direction r sha256 f99121055e9090a5a85d39002fe3e83f2d77119b1943f494497cf262c5a6f4ad (abliteration.json in this repo); 18 shards LFS-verified on download
DFlash2 drafter	z-lab/Qwen3.8-27B-DFlash2	z-lab	50307d4c4cde6860d4eee73e2547cd786fe8e8a4
Engine	Neroued/ninfer	Neroued	v3 container, ≥ 98dada0e
Quantization-runbook inspiration	Barding-Defense/Qwen3.8-27B-huihui-abliterated-NVFP4-NInfer	Barding-Defense	community ninferization walkthrough

What we did

1. Abliteration — single-direction Arditi-style ablation: orcarouter's refusal direction r, recovered from their Qwen3.8-27B-Uncensored weights, projected out of 131 residual-writing tensors of Swift-1.5 (o_proj, GDN out_proj, mlp.down_proj, embed_tokens) in float32, exactly as ajgazin's release.
2. All-NVFP4 quantization with llm-compressor (512 Ultrachat calibration samples, seq 2048, sequential CPU pipeline): NVFP4 W4A4 group-16 on every text projection, W8G32 token embedding + output head, official q6/q8 vision allocation, MTP + DFlash2 BF16.
3. Global-divisor normalization across each fused packing group (worst re-encode rel err 79.25%), then v3 conversion with the fork converter (--components text,vision,mtp,dflash2 --proposal).

Measured results (same-session A/B on one RTX 5090, temp 0)

metric	Swift-1.5 (prod, censored)	this artifact
Refusal rate (harmful_behaviors test[:100], greedy)	59/100	8/100 (harmless sanity 45/50)
Harmless sanity pass rate (n=50)	43/50	45/50
IFBench prompt-strict (n=300, fresh)	69.0	70.3 (loose 74.3, instr 70.6/74.1)
GSM8K-200 (fresh)	95.0%	94.5% (−0.5 pp, n=200 noise)
Gate decode tok/s	145.5	163.2 (GATE PASS)
Perf decode (mean of 3)	167.4	165.7 (parity)
Prefill @ 200k	3,259 tok/s	3,259 tok/s (parity)
llama-benchy tg @ depth 0/16k	202.6 / 189.2	198.9 / 211.9 (full grid below)
Needle full ladder (haystack 1M chars / 250,031 tok, 3 depths)	EXACT	24/24 EXACT, 0 corrupted

Artifact

Field	Value
Filename	qwen3_8_27b_swift15abl_nvfp4full-dflash2.ninfer
Size	19,782,447,364 bytes (18.42 GiB)
SHA-256	65f9d2ab1161e95d65ecdb20858b42ae937bb40c92d3dabf969fd47bfd03c0cb
Container version	3
Formats	nvfp4 ×256 (fused text parents), q8_g32 ×30, q4/q5/q6 (vision), bf16 remainder

Serving (RTX 5090 32 GB, single GPU)

ninfer-serve qwen3_8_27b_swift15abl_nvfp4full-dflash2.ninfer \
  --model-id Qwen3.8-27B --max-context 262144 --kv-capacity auto --kv-dtype k8v4 \
  --max-concurrency 4 --default-max-tokens 32768 --prefill-chunk 4096 \
  --temperature 0.9 --min-p 0.05 --spec dflash2 --draft-tokens 7 --lm-head-draft \
  --host-kv-mib 49152 --host-state-slots 16 --vision \
  --default-thinking-budget 16384 --preserve-thinking --image-token-budget 1280

---

Acceptable Use Policy & Legal Notice

By downloading, possessing, or using this artifact you agree to the terms below.

Permitted use

Research, personal/local experimentation, red-teaming, safety research, and evaluating alignment-removal techniques — subject to all applicable laws and regulations.

Prohibited use

You must not use this model, alone or in any pipeline, to:

• carry out, plan, or facilitate any activity that is illegal in your jurisdiction;
• produce content that harms, endangers, defrauds, harasses, or exploits others — including but not limited to weapons, malware, exploitation of minors, targeted harassment, or disinformation presented as fact;
• provide medical, legal, or financial advice presented as professionally qualified;
• violate the rights of any person or entity, including intellectual property and privacy rights.

Your responsibility

This model is a capability, not a judgment. Abliteration removes refusal behavior; it does not make outputs trustworthy, factual, or legal. You are solely and fully responsible for every prompt you send, every output you generate, and every use you make of them. The publisher of this artifact (kaushikvira) does not monitor, endorse, or take any part in downstream use.

No warranty / limitation of liability

The artifact is provided "AS IS", WITHOUT WARRANTY OF ANY KIND, express or implied, including merchantability, fitness for a particular purpose, and non-infringement. To the maximum extent permitted by applicable law, the publisher shall not be liable for any claim, damages, or other liability, whether in contract, tort or otherwise, arising from, out of, or in connection with this artifact or its use. Nothing in this card limits liability where limitation is not permitted by law.

Measurement conditions

All numbers on this card were measured on a single RTX 5090 32 GB (driver 595.71.05):

• GPU power cap 450 W (stock 575 W) and SM clock pinned 2280 MHz via the nv-power-limit.service systemd unit — a thermal-efficiency mod (load power ~450 W → ~330 W) with no measurable tok/s loss vs stock.
• Telemetry correlated with the runs below (70 s thermal-log samples plus a 2 s-sampled instrumented run): peak board draw 375–433 W (cap 450 W), SM clock 2248–2272 MHz (pin 2280), GPU temp 36–69 °C, util 100 % under load.
• llama-benchy 0.4.0 against the engine's native OpenAI API (not llama.cpp), greedy, mean ± std of 3 runs; tokenizer ukisai/Swift-1.5-Qwen3.8-27b.
• pp8192 @ dN = throughput of 8,192 new tokens processed on top of N cached context tokens — the agentic-coding profile (long, growing context; small incremental prompts). e2e ttft is end-to-end time to first response chunk.

llama-benchy (single RTX 5090, engine-native API)

test	t/s (mean ± std of 3)
pp2048	48,625 ± 35
tg256	199 ± 18
pp2048 @ 16k cached ctx	387,707 ± 2,966
tg256 @ 16k cached ctx	212 ± 17
pp8192 @ 32k cached ctx	746,777 ± 2,595
tg512 @ 32k cached ctx	184 ± 18
pp8192 @ 131k cached ctx	1,428,495 ± 13,022
tg512 @ 131k cached ctx	188 ± 26
pp8192 @ 200k cached ctx	1,729,533 ± 11,547
tg512 @ 200k cached ctx	167 ± 9

kaushikvira Qwen3.8-27B artifact family (single-RTX-5090 serving)

All artifacts are NInfer containers of Qwen3.8-27B derivatives, served by the same engine on the same box. Per-card A/B tables were measured pairwise in the same session; cross-card rows are from different sessions — treat small deltas as indicative.

artifact	base	uncensored	IFBench prompt-strict	GSM8K-200	decode tok/s	tg512 @ 131k ctx	status
nvfp4full-dflash2 (v2)	Qwen3.8-27B	no	—	—	146.4*	—	superseded by v3 (same weights, v2 container)
nvfp4full-dflash2.v3	Qwen3.8-27B	no	65.0*	96.5%*	146.4*	162.9	available
swift-abliterated v3	d0xin Swift-1.0 (huihui-abliterated)	yes	66.3	95.5%	148.7	173.7	available
swift15 v3	ukisai Swift-1.5	no	69.0	95.0%	160.8	153.4	rollback profile
thinkingcap v3	BottleCapAI ThinkingCap	no	68.0	95.5%	167.6	177.3	available (PolyForm — non-commercial)
swift15-uncensored-ajgazin v3	ukisai Swift-1.5 + ajgazin/orcarouter ablation	yes	70.3	94.5%	165.7	187.9	current production

* values for the nvfp4full-dflash2.v3 row come from its same-session A/B against swift-abliterated (66.3-vs-65.0 etc. are paired measurements, not independent runs).

HF-format variant (vLLM / SGLang — JSON-schema structured output)

NInfer does not support JSON-schema output; if you need structured output or the standard vLLM/SGLang stack, use the HF-format NVFP4 weights:

• ajgazin/Swift-1.5-Qwen3.8-27B-Uncensored-NVFP4 (official, by the abliteration author).

If you benchmark an HF-format variant on your hardware, please share numbers in the repo's Community tab — we collect them on the variant's card.

License terms carried by this artifact

• Swift Contribution (ukisai): Swift Open License v1.0 — a full copy is in LICENSE. Note its Commercial Use Limitation (§5): commercial use by entities over US$1M gross annual revenue requires a separate enterprise license from UkisAI. This limitation applies to the Swift Contribution contained in this derivative.
• Base Model (Qwen3.8-27B): Apache License 2.0 — copy in LICENSE-APACHE-2.0.
• Modified files notice (Swift Open License §4b): the weights in this repository are modified relative to the upstream Work — refusal-direction ablation (131 tensors, listed in abliteration.json) plus NVFP4/q8 requantization and NInfer v3 containerization.
• NOTICE: attribution notices preserved in NOTICE.

Reporting

If you become aware of misuse of this model, report it to the platform where the misuse occurs and to the relevant authorities. Do not contact the publisher for takedowns of third-party actions — the publisher has no control over downstream deployments.