← back to catalog · registered 2026-11-18 15:04

Austriani/TwIL-8B-ARA-Abliterated

Austriani 8B GGUF second-order
Your rig guess connected
? Why do I need an app?
Reading your rig…

This is a rough estimate. Install the free app - we'll show exact numbers.

Reading real hardware from your app right now. Numbers below are exact.

Below is the per-quantization compatibility for this model.

curl -H "Authorization: Bearer $ABL_KEY" \
     "https://abliteration.org/api/v1/models/Austriani%2FTwIL-8B-ARA-Abliterated"
Response includes
  • classification m8
  • files 2
  • author_summary 13 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M8
Primary method

Repackaging (quantization)

Applied on top of direct removal inherited from the base model.
Confidence
MEDIUM
Inherited from base model
Why this label 3 signals
Method inferred from partial signals - repository name, related files, or tag patterns. Producer identity not confirmed; label may sharpen or shift as we gather more evidence.
  • 'abliterated' in name/tags
  • is_gguf=1
  • assume M1 (base ablation) + M8 (GGUF quant) - default when producer unknown
Refusal direction extraction

No specific extraction method could be identified for this model. The producer either did not document it or used a proprietary pipeline.

What is a refusal direction? →
Downloads · 30-day
565
Likes
0
Descendants
2
in 2 direct forks
Model age
3d ago
created 2026-11-18
Downloads over time
Now 565 → from 0 ↑ 0%
0207414622 0 on Nov 18 565 on Nov 22 Nov
Nov 18 → Nov 22 · 5 snapshots · spans 4 days

Genealogy 2 direct forks

Full fork graph →

This model's place in the market. Above: what it was derived from. Below: the tree of everything derived from it.

Variants by this author 2 formats · 2K downloads combined

The same weights this author released in different packaging. Pick the format that matches your runtime.

Metadata

License
other
Languages
en
Tags
transformers safetensors granite text-generation twil twil-8b granite-4.2 8b reasoning math code instruction-following

Related

Total size
0 B
Files
2
Quantizations
1
Registered
2026-11-18 15:04
Last updated on HF
2026-11-18 18:50

Files by quantization

Auxiliary files 2 files 17.1 KB
README.md 15.6 KB 5e81f4d0 download
.gitattributes 1.48 KB a6344aac download

README current version from Hugging Face


base_model: Austriani/TwIL-8B-ARA-Abliterated
pipeline_tag: text-generation
library_name: ik_llama.cpp
license: other
license_name: webai-non-commercial-license-ver.-1.0
license_link: https://huggingface.co/webAI-Official/TwIL-8B/blob/main/LICENSE.md
language:

  • en
    tags:
  • twil
  • twil-8b
  • granite
  • granite-4.2
  • 8b
  • reasoning
  • math
  • code
  • instruction-following
  • abliterated
  • uncensored
  • nsfw
  • ara
  • arbitrary-rank-ablation
  • heretic
  • refusal-removed
  • ik-llama.cpp
  • iqk
  • trellis-quant
  • maxoptima
  • gguf
  • quantized
  • imatrix
  • v3-calibration
  • austriani
  • webai

Austriani // ik_llama.cpp // GGUF

TwIL 8B
ARA Abliterated

Granite 4.2 Dense // IQK Quantization // V3 Calibration
8.79B
Parameters
40
Layers
131K
Context
1024
Imatrix Ctx
The Base Model
This repository hosts ik_llama.cpp-compatible GGUF quantizations of Austriani/TwIL-8B-ARA-Abliterated. That checkpoint is an Arbitrary-Rank Ablation of webAI-Official/TwIL-8B — specifically the verbose variant, the RL-stage output with longer reasoning traces.
TwIL-8B is webAI's 8.79B reasoning model built on IBM Granite 4.2 8B via full-parameter SFT followed by GRPO/DAPO reinforcement learning. It is a 40-layer dense decoder (GraniteForCausalLM) with hidden size 4,096, 32 attention heads and 8 KV heads, a 100,352-token vocabulary, and a 131,072-token context window inherited from Granite. It emits a <think> block before its answer by default.
Parent-reported strengths: MATH-500 98.2, LiveCodeBench 81.6 (verbose), IFBench 74.4 (verbose), AIME 2025 88.3. The model is a reasoning specialist, not a knowledge model — GPQA Diamond sits at 70.2, below Qwen3.5-9B's 81.7. It uses roughly 38% fewer tokens than Qwen3.5-9B on AIME 2026 while landing in the same accuracy band. These are upstream numbers; validate on your own workloads.
License note: The parent TwIL-8B ships under the webAI Non-Commercial License 1.0, not the Apache 2.0 of its Granite base. The abliterated checkpoint and these quants inherit that restriction. The nsfw tag is applied because the ARA ablation removes refusal behaviour — this is not a safety-tuned model.
Abliteration Report // ARA Method
The base abliteration uses Arbitrary-Rank Ablation (ARA) via Heretic, applied across the full 40-layer stack (0–39). The optimizer settled on trial 25 of 40, with a measured refusal rate of 14/100 and a KL divergence of 0.0412. The preserve weight of 0.7003 is high relative to the steer weight (0.0002), which means the optimizer prioritized keeping harmless behaviour and general capability intact while the overcorrect weight of 0.8309 carried the decensoring pressure.
Run metrics, not guarantees. A 14% refusal rate means most refusals are gone, but not all. A KL of 0.0412 is low for an ARA run — closer to the parent distribution than many abliterations — so reasoning and coding capability should survive better than a heavier ablation would allow. Still, validate on your own tasks.
View Full ARA Trial Parameters
trial
25 / 40
start_layer_index
0
end_layer_index
39
preserve_good_behavior_weight
0.7003
steer_bad_behavior_weight
0.0002
overcorrect_relative_weight
0.8309
neighbor_count
2
refusals
14 / 100
KL divergence
0.0412
MaxOptima Custom Quant
This release includes MaxOptima, a layer-aware mix for the 40-layer Granite dense stack. The strategy is boundary-weighted: embeddings and output at Q6_K, attention V at IQ6_K everywhere, attention K lifted on every sixth layer, FFN-down protected at IQ5_KS on boundary and mid layers, and the first fifteen layers compress gate/up more aggressively than the rest. Layer 64 is pinned to Q4_K as a stabiliser.
MaxOptima strategy: Embeddings and output locked at Q6_K. Layer 64 forced to Q4_K. Attention V at IQ6_K globally, with layers 0 and 58 explicitly IQ6_K. Attention K at IQ6_K on every sixth layer (5, 11, 17, 23, 29, 35, 41, 47, 53, 59). FFN down at IQ5_KS on boundary and mid layers (0, 5, 11, 17, 23, 25, 26, 29, 35, 41, 47, 53, 59), IQ4_K elsewhere. FFN gate/up at IQ4_KS on layers 0–14 and IQ4_K on layers 50–59. Attention output at IQ6_K on layer 59. Default fallback IQ4_KS.
View Full MaxOptima Custom Quant Configuration
^output\.weight=Q6_K,token_embd\.weight=Q6_K,blk\.64\..*=Q4_K,blk\.(0|58)\.attn_v\.weight=IQ6_K,attn_v\.weight=IQ6_K,blk\.(5|11|17|23|29|35|41|47|53|59)\.attn_k\.weight=IQ6_K,blk\.(0|5|11|17|23|25|26|29|35|41|47|53|59)\.ffn_down\.weight=IQ5_KS,blk\.(0|1|2|3|4|5|6|7|8|9|10|11|12|13|14)\.ffn_(gate|up)\.weight=IQ4_KS,blk\.(50|51|52|53|54|55|56|57|58|59)\.ffn_(gate|up)\.weight=IQ4_K,blk\.59\.attn_output\.weight=IQ6_K,ffn_down\.weight=IQ4_K,.*=IQ4_KS
Available Quants
MaxOptima // Custom Mix
MaxOptima
Custom layer-aware mix. Q6_K embeddings/output, IQ6_K attention V, IQ6_K attention K on every sixth layer, IQ5_KS FFN-down on boundary and mid layers, IQ4_KS/IQ4_K FFN gates, Q4_K pinned boundary block. Built for the 40-layer Granite dense stack.
Standard IQK // High & Mid Fidelity
IQ5_K
Higher fidelity 5-bit. Best quality-to-size for systems with sufficient memory.
IQ4_K
Standard 4-bit IQK. Uses 256-value super-blocks for scaling.
IQ4_KS
"Small" 4-bit variant with tighter packing. Lower memory footprint.
IQ4_KS_R4
Row-interleaved variant of IQ4_KS, optimized for ik_llama.cpp's matmul kernels.
Trellis // Alternative 4-Bit
IQ4_KT
Dynamic programming trellis quant. Uses a fundamentally different error-minimization strategy than IQK.
Imatrix Specification
This release uses the V3 Calibration Dataset used across previous Austriani releases. Calibration ran at 500 cycles per chunk with a 1024-token context window, processing approximately 1000 chunks.
Companion Repositories
These GGUF files are built exclusively for ik_llama.cpp. IQK, Trellis, and MaxOptima quants will not load in mainline llama.cpp. For LM Studio, Ollama, KoboldCPP, or stock llama.cpp, use the baseline GGUF repository above.

README history 6 versions

The author's README evolved over time. Click a version to see its content at that point.

  1. 2026-11-18 Update README.md 068a7cd 12.5 KB
    Loading...
  2. 2026-11-18 Update README.md 62e6067 15.6 KB
    Loading...
  3. 2026-11-18 Update README.md 121f04d 15.6 KB
    Loading...
  4. 2026-11-18 Update README.md 1e1e71b 15.5 KB
    Loading...
  5. 2026-11-18 Update README.md 78b462d 15.5 KB
    Loading...
  6. 2026-11-18 Create README.md 9381c21 15.5 KB
    Loading...
Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Open in Abliteration