Barding-Defense/Qwen3.8-27B-OBLITERATED-NVFP4-NInfer
Qwen3.8-27B OBLITERATED — NVFP4 for NInfer
An NVFP4 artifact of OBLITERATUS/Qwen3.8-27B-OBLITERATED, packaged for the NInfer engine. One file carries the complete container — Text, Vision, MTP, the optimized proposal head, tokenizer, chat template, generation config and media-processor resources.
qwen3_8_27b_obliteratus_nvfp4.ninfer
21,492,695,040 bytes
sha256 f6f404bd33fe125830adbdd47434860f77a3ba97202086280bb4f4f888eb3902
identity qwen3.8-27b / nvfp4 container v2 1124 objects⚠️ The frontend resources were replaced — read this
The NInfer converter gates on SHA-256 of six frontend files. This source failed that gate 3/6, so `tokenizer_config.json`, `chat_template.jinja` and `generation_config.json` were replaced with the canonical `Qwen/Qwen3.8-27B` versions. Full record in frontend-substitution.json.
This was a repair, not an override. The source's chat template is 506 bytes with no vision handling at all — no image/video macros, no <|vision_start|> tokens — a text-only template on a multimodal model, which would have broken image input entirely. Its generation_config.json had also dropped an EOS token (248044), risking runaway generation.
Weights are untouched. Only the frontend resources differ from what OBLITERATUS published. If you specifically want their chat template, this artifact is not what you want.
Why this exists
Qwen3.8-27B is the king of the local models right now, and the great work Neroued has done with the NInfer engine runs it extremely quickly on Blackwell. There has also been a lot of work on obliterating and otherwise liberating models, including Pliny's OBLITERATUS suite.
Some abliterated models have been "ninferized" already — see Related work below. Two things I hadn't seen:
- the walkthrough/recipe to do it yourself
- Pliny's OBLITERATUS Qwen3.8-27B, ninferized
Both are offered here. The recipe lives in the NVFP4 repo — AGENT.md, SUPPORTED.md and scripts/. Under $5 of RunPod time converts a compatible checkpoint; measured runs came in at $1.69 and $3.30.
What it is
Format conversion only. No fine-tuning, no re-abliteration, no behavioural editing of the weights. Refusal characteristics are inherited from the upstream OBLITERATUS checkpoint.
Numeric allocation
Structurally identical to the official NVFP4 artifact — same 1124 objects, same per-format and per-layout counts, same file size to the byte. Only values differ.
Requirements
- Blackwell GPU, compute capability 12.0. Verified on an RTX PRO 6000 Blackwell; an
sm_120abuild runs on any cc-12.0 device, not only the RTX 5090. - ~19.0 GiB VRAM for weights.
- Stock NInfer — no engine patch required.
Quick start
ninfer qwen3_8_27b_obliteratus_nvfp4.ninfer \
--prompt "..." --max-context 16384 --max-new 256 \
--spec mtp --draft-tokens 5 --lm-head-draftninfer-serve qwen3_8_27b_obliteratus_nvfp4.ninfer \
--model-id qwen3.8-27b-obliteratus \
--max-context 262144 --kv-capacity auto --max-concurrency 1 \
--kv-dtype nvfp4 --spec mtp --draft-tokens 5 --lm-head-draft \
--preserve-thinking --vision --host 0.0.0.0 --port 8080Measured
Verified on an RTX PRO 6000 Blackwell with an engine built from source at a140e7ae.
Zero fallbacks means the MTP head survived conversion intact.
For comparison, the huihui NVFP4 build measured 88.8 % acceptance under identical flags. Both are healthy; the spread is a genuine difference between source checkpoints, plausibly because OBLITERATUS ablated more aggressively relative to the unmodified MTP head. Reported rather than smoothed over.
Refusal behaviour — measured on this artifact
AdvBench harmful_behaviors (Zou et al. 2023), n = 20
refuse 0 0.0%
comply 20 100.0%
incoherent 0 0.0%Classifier: the refusal-prefix heuristic standard in the abliteration literature (Arditi et al. 2024 family). Full record in refusal-check.json; it stores verdicts and refusal wording only, not complying completions.
Caveat: no baseline was measured. Confirming that stock Qwen3.8-27B refuses these same prompts would require downloading another 55 GB, which was not done. So this says "complies with 20/20 AdvBench prompts a safety-tuned model declines" — strong evidence abliteration survived quantization and conversion, but not a paired before/after comparison.
Scope of validation
Verified: loads on stock NInfer; MTP engages at 77.8 % with zero fallbacks; thinking mode works; coherent, correct output on reasoning, code and explanation prompts; container structurally equivalent to the official artifact; source intact (1199 tensors, 15 MTP, 333 vision, all BF16, shapes matching official); refusal rate 0/20 as above.
Not run: no capability benchmark and no fidelity comparison against BF16. No HumanEval+/GPQA numbers are claimed because none were produced. Abliteration has a documented capability cost and this artifact inherits it, unquantified.
Verify
sha256sum -c SHA256SUMS
python3 -m tools.artifact.inspect qwen3_8_27b_obliteratus_nvfp4.ninferExpect identity qwen3.8-27b/nvfp4, objects 1124 (1118 tensors, 6 resources), file_bytes 21492695040.
Reproducing
The full agent-executable procedure, all scripts, and every gotcha are in the huihui NVFP4 repo: AGENT.md, SUPPORTED.md, REPLICATION.md and scripts/. This artifact was produced by following that procedure end to end on a single rented Blackwell pod (~95 min, ~$3.30), including the frontend substitution branch documented above.
Related work — other NInfer abliterated artifacts
Other people got here first, and their artifacts are also good. Use whichever fits — this list is here so you can find them.
Worth noting three are built on `orcarouter/Qwen3.8-27B-Uncensored`, which is gated.
As far as I can tell, none of the above ship a conversion recipe. That is what this repo adds.
Licence
Apache-2.0, inherited from the chain: Qwen/Qwen3.8-27B (Copyright 2026 Alibaba Cloud), OBLITERATUS/Qwen3.8-27B-OBLITERATED, Neroued/ninfer and unsloth/Qwen3.8-27B-NVFP4 are all Apache-2.0. LICENSE is the canonical file with its copyright notice intact (§4(c)); NOTICE carries the attribution chain and statement of changes (§4(b)). Calibration used HuggingFaceH4/ultrachat_200k (MIT); it affects only per-matrix input_global_scale scalars and reproduces no dataset content.
"Qwen" is a trademark of Alibaba Cloud. Unofficial community derivative, not endorsed by or affiliated with Alibaba Cloud, OBLITERATUS, unsloth, or the NInfer project.
A description of what the licences say, not legal advice.
Intended use and limitations
This is an uncensored model with a measured 0% refusal rate on AdvBench. It will attempt requests a safety-tuned model declines. Published for research, evaluation and local deployment where that is understood and wanted. Anyone deploying it is responsible for their own safeguards and for compliance with the Apache-2.0 licence and applicable law. Refusal behaviour and any capability regression originate upstream — this repository changes format, not weights.
Credit for the abliteration to OBLITERATUS; the base model to Qwen; the engine and artifact contract to @Neroued; the NVFP4 recipe to unsloth.