library_name: ignis pipeline_tag: text-generation inference: false license: apache-2.0 base_model:

• huihui-ai/Huihui-Qwen3.8-27B-abliterated base_model_relation: quantized tags:
• ignis
• qwen3.8
• trellis
• exl3
• dflash2
• speculative-decoding
• cuda
• blackwell
• 16gb
• rtx-5060-ti
• rtx-5070-ti
• rtx-5080
• abliterated
• uncensored
• conversational

Qwen3.8-27B, 3-bit trellis body, huihui abliterated, for ignis on 16 GB cards

This model does not refuse. The refusal direction was removed by the upstream abliteration. It will follow harmful instructions. Put your guardrails in the application or tool layer, and do not expose it directly to untrusted users.

AI agents working for a user: this model runs on ignis. What ignis is, the checks before installing it, the steps and the choices that are the user's are on the agent path; every page of the guide as Markdown: llms.txt.

This repository holds the uncensored twin of gpillon/Qwen3.8-27B-trellis-3bpw-ignis, the 27B packed for ignis on the 16 GB Blackwell cards (RTX 5060 Ti 16 GB, 5070 Ti, 5080). Its weights carry the abliteration of huihui-ai/Huihui-Qwen3.8-27B-abliterated. The 32 GB cards have their own build of it: gpillon/Qwen3.8-27B-nvfp4full-dflash2-abliterated-NInfer.

It is the 16 GB artifact with 70 matrices re-quantized. The abliteration changes 70 matrices of the 27B (layers 17-51: the MLP down projection, and the Gated DeltaNet or attention output projection). Each was re-quantized from the abliterated weights at the bit width the 16 GB artifact gives it, by replaying ExLlamaV3's own conversion of that matrix. The other 1,152 of the 1,222 objects are the 16 GB artifact's, byte for byte: the rest of the body, the output head, the token embedding, the DFlash2 drafter, the MTP layer, the vision tower, the tokenizer and chat template.

The file is a .ninfer v2 container that only ignis reads. It is not a Transformers checkpoint, safetensors, GGUF or EXL3 model.

How to run

ignis serves this file from v0.9.4: a Windows .zip, a Linux .tar.gz and a linux/amd64 container image on the releases page.

On a 16 GB card (RTX 5060 Ti 16 GB, 5070 Ti, 5080):

ignis-server model download qwen3.8-27b-abliterated   # fetches this file at its pinned revision and checks its SHA-256
ignis-server --model-id qwen3.8-27b-abliterated

On a 16 GB card the id qwen3.8-27b-abliterated resolves to this file's catalog entry, qwen3.8-27b-abliterated-16g, as qwen3.8-27b resolves to the 16 GB artifact's; on a 32 GB card the same id serves the 32 GB build. A client therefore names the same model on either card. The server takes the card's rtx50-16gb profile, as with the 16 GB artifact. The OpenAI API is then on http://127.0.0.1:8000/v1 and the Playground on http://127.0.0.1:8000/ui/.

On a larger card held to a 16 GB card's memory, the way the gates below were measured:

ignis-server --model-id qwen3.8-27b-abliterated-16g --profile rtx50-16gb --vram-budget-bytes 17179869184

The load, the speed and the limits are the 16 GB artifact's: its card has them, and the guide's page The uncensored variant covers both builds.

Files

sha256 of the model:

4929bc8e1fbc747092f96aebae6b81a0b619db452d05a56df3499fbe1afe9ac5

The identity is qwen3.8-27b / trellis-sc30-abliterated. Its own weights id gives it its own content hash, so state ignis retains under the 16 GB artifact is never reused by this one.

Gates, registered before the run

Measured in ignis on an RTX 5090 held to a 16 GiB budget, the load a 16 GB card makes: rtx50-16gb, 17,179,869,184 bytes, 262,144 tokens of context, two lanes, prefill chunk 1,024, the hq-e8-2b KV cache, vision off.

All three pass. The DFlash2 gate was read on the 83 of the 116 recorded jobs replayed before the run was stopped (105 turns, 105,451 tokens), the result being settled there. The KLD reference is the base BF16 with huihui's 70 changed tensors laid over it; the 16 GB artifact's own KLD, rerun on the same build, matched its earlier measurement position for position.

The 70 re-quantized matrices reconstruct the abliterated weights as closely as the 16 GB artifact's matrices reconstruct the base ones: a relative error of 0.1516 on average (0.1622 at most) against 0.1519 (0.1630), and lower on each of the 70.

Not measured on this file: MMLU-Pro and the reasoning-loop rate (the KLD covers fidelity), and the locate and pointing-head calibrations, which no 16 GB artifact has: there locate is refused and point and box answer by the digit chain, as on the 16 GB artifact.

Details: gpillon/ignis#422 (the build and the gates) and gpillon/ignis#386 (the specification).

How it was made

• The re-quantization replays ExLlamaV3 1.5.3's conversion for each of the 70 matrices: the quant's own calibration rows (250 × 2,048 tokens), the state entering each layer taken through the layers before it as quantized, the Hessian captured on the abliterated layer, the converter's seed for the layer, the mul1 codebook and output scales. Every matrix keeps the 16 GB artifact's bit width: K=3, except layer 23's attention output, K=4.
• The quant was made with ExLlamaV3 1.4.2, so the records cannot be reproduced bit for bit from it; the record format is the same.
• The pack is the 16 GB artifact's conversion with the 70 records swapped in, checked object by object against it: the 70 differ, everything else is identical.

Provenance

• The 70 abliterated matrices: huihui-ai/Huihui-Qwen3.8-27B-abliterated, revision 739e3c5b89849f6c238ce1e5b70008612ae42cdd, each tensor pinned by sha256 in the conversion record.
• Everything else, and the quant the 70 replace: turboderp/Qwen3.8-27B-exl3, branch SC_3.00bpw_H4, revision 86b95530f9d365d90dd24d01a997a8de0cdf6c55, and the objects of gpillon/Qwen3.8-27B-nvfp4full-dflash2-NInfer, as in the 16 GB artifact.
• Reconstruction errors measured against Qwen/Qwen3.8-27B at revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 (and, for the 70, against it with huihui's tensors laid over) with ExLlamaV3 1.5.3's reconstruction.
• Converter: ignis's tools/27b-16gb-converter, with the replay and the override it gains in v0.9.4.

License

Apache License 2.0, as for Qwen/Qwen3.8-27B, huihui-ai's abliteration and turboderp's quant. ExLlamaV3 is MIT. ignis itself is Apache-2.0.