For privacy reasons a browser tells us at most
"≥ 8 GB RAM, 8 cores" - same reading whether you have 8 GB
or 128 GB. It has no idea how much RAM is free right now, which
apps are open, or whether you have a GPU.
The Abliteration app is integrated with your machine
It reads your exact RAM, GPU model and VRAM,
free memory right now, and picks the sharpest quant that
still fits. Every model page lights up precisely for your rig.
And you can chat with any model, right now
The app is a full local runtime - no API keys, no subscription,
everything runs on your machine. Click any model on this site and
start a conversation in seconds.
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=0 (base model)
no specific method indicators - defaulting to M1 (most common)
Qwen3.5 Hybrid Decoder // 262K Context // Agentic Base
9B
Parameters
262K
Native Context
42%
Refusal Rate
0.3377
KL Divergence
The Base Model
This repository hosts the ARA-abliterated variant of XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B. The parent is a 9-billion-parameter agentic model built by Xiaomi MiMo via supervised fine-tuning of Qwen3.5-9B on MiMo-generated data spanning software engineering, terminal tool-use, visual coding, and cybersecurity.
Architecturally, it inherits the Qwen3.5 hybrid attention stack: 24 gated DeltaNet linear-attention layers interleaved with 8 full-attention layers across 32 blocks (full attention every 4th layer). The linear layers carry a fixed-size recurrent state instead of a growing KV cache, which keeps memory growth flat at long context. A vision encoder feeds image tokens into the same text backbone — text and image input share one model, not an adapter.
Parent training notes: The SFT mixture totals 77.4B tokens (27.2B loss-bearing), weighted toward code (29.9%), general agent tasks (28.5%), visual (27.4%), and cyber (14.2%). Thinking mode is exposed through an enable_thinking switch in the chat template, with reasoning content returned separately from the final answer.
Abliteration Report // ARA Method
The abliteration uses Arbitrary-Rank Ablation (ARA) via Heretic. The optimizer balances three competing objectives: preserve harmless behavior, steer harmful outputs toward harmless ones, and overcorrect harmful outputs away from their original state. The run settled on trial 18 of 40, with a measured refusal rate of 42/100 and a KL divergence of 0.3377.
Run metrics, not guarantees. A 42% refusal rate means refusals are reduced, not eliminated. KL divergence of 0.3377 is moderate — capability retention should be validated on your own tasks. This is an SFT checkpoint intended as a starting point for agentic RL research; the abliteration does not change that.
View Full ARA Parameters
start_layer_index
7
end_layer_index
29
preserve_good_behavior_weight
0.5826
steer_bad_behavior_weight
0.0465
overcorrect_relative_weight
0.8496
neighbor_count
9
trial
18 / 40
refusals
42 / 100
KL divergence
0.3377
What Changed
The ablation targets layers 7 through 29 — the middle-to-late blocks where refusal directions are most concentrated in the Qwen3.5 hybrid stack. The preserve weight of 0.5826 is high relative to the steer weight (0.0465), which means the optimizer prioritized keeping harmless behavior intact over aggressively flipping harmful outputs. The overcorrect weight of 0.8496 is where the actual decensoring pressure sits.
Because this is a hybrid linear/full attention model, the abliteration operates on a different geometry than a standard transformer. The linear-attention layers hold recurrent state rather than explicit KV pairs, so the refusal direction lives in the state transition weights, not in attention outputs. The ARA implementation accounts for this by ablating across the full block span rather than per-attention-head.
Usage Notes
This is a research artifact. It inherits the parent's thinking mode, tool-use formatting, and 262K context window. It also inherits the parent's vision encoder — image input works. The abliteration was performed on the text backbone; the vision tower is untouched.
The refusal rate of 42% means some prompts will still be declined. If you need a cleaner refusal surface, the parent model card links to additional abliteration trials with different trade-offs between refusal reduction and KL divergence. This trial was selected for the middle of that curve.
Format: BF16 safetensors. Load with transformers (Qwen3_5ForConditionalGeneration). The chat template and tokenizer are included. For GGUF conversion, the hybrid architecture requires a recent llama.cpp build with Qwen3.5 delta-net support — older builds will fail on the linear-attention blocks.
README history
2 versions
The author's README evolved over time. Click a version to see its content at that point.
2026-11-13Update README.md44b92db11.9 KB
Loading...
2026-11-13initial commit062574a28 B
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.