---
library_name: transformers
license: apache-2.0
license_link: https://huggingface.co/Qwen/Qwen3.6-35B-A3B/blob/995ad96eacd98c81ed38be0c5b274b04031597b0/LICENSE
base_model: Qwen/Qwen3.6-35B-A3B
pipeline_tag: image-text-to-text
tags:
  - qwen3.6
  - moe
  - lora
  - merged
  - antidoom
---

# Qwen3.6-35B-A3B-AntiLoop

This is a BF16 merged-weight variant of
[`Qwen/Qwen3.6-35B-A3B`](https://huggingface.co/Qwen/Qwen3.6-35B-A3B),
fine-tuned to recover from pathological self-verification and enumeration loops
while preserving ordinary long-form reasoning.

The artifact contains a full Transformers checkpoint; no PEFT adapter is needed
at inference time. It retains the base model's multimodal architecture,
tokenizer, chat template, MTP weights, and 262,144-token native context length.

## Provenance

- Base: `Qwen/Qwen3.6-35B-A3B`
- Pinned base revision: `995ad96eacd98c81ed38be0c5b274b04031597b0`
- Fine-tuning: rank-32 LoRA, alpha 64, dropout 0
- Targets: 190 attention-side projections across all 40 language-model layers
  (150 Gated DeltaNet projections and 40 full-attention projections)
- Objective: `0.9 * supervised CE + 0.1 * KL(base || tuned)`
- Training: one epoch / 178 optimizer steps, seed 0
- Training run: `n8programs/qwen-antidoom`, run `dw7mbmce`

The LoRA was merged as
`BF16(W_FP32 + B_FP32 @ (A_FP32 * 2))`, one target tensor at a time in
small row tiles. `merge_manifest.json` records the exact adapter hash, base
revision, environment, per-target effective deltas, and every output-shard
SHA-256. `merge_verification.json` certifies that all 190 intended tensors equal
that formula bit-for-bit after BF16 rounding, every target differs from the
base, and every non-target byte is unchanged.

## Evaluation

The behavior evaluations below used the same round-2 adapter served on the
Qwen3.6-35B-A3B FP8 base with vLLM. They validate the adapter behavior; the
merged BF16 artifact was separately validated at the tensor level.

| Evaluation | Base/control | AntiLoop |
|---|---:|---:|
| Held-out enumeration-loop rate (285 prompts, calibrated judge) | 25% | **2%** |
| Answer delivery on the same prompts | 77% | **98%** |
| GPQA Diamond, thinking enabled | 167/198 (84.3%) | **166/198 (83.8%)** |

The paired GPQA difference was -0.5 percentage points (exact two-sided McNemar
`p = 1.0`), consistent with preserved capability at this sample size. The loop
evaluation is task-specific and judge-based; it should not be read as a general
safety or truthfulness guarantee.

## Usage

Use a recent Transformers release with Qwen3.6 support:

```python
from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "N8Programs/Qwen3.6-35B-A3B-AntiLoop"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    dtype="auto",
    device_map="auto",
)
```

The checkpoint can also be served directly with recent vLLM/SGLang builds that
support the Qwen3.6 MoE architecture. Follow the upstream Qwen3.6 model card for
chat templating, thinking-mode controls, multimodal inputs, deployment details,
and the base model's limitations.

## Limitations

- This is a narrow behavioral fine-tune, not a general alignment or safety model.
- The merged BF16 checkpoint has not been exhaustively evaluated across every
  language, modality, tool-use setting, or long-context regime.
- Quantization can change behavior; validate the exact deployed quantization.
- Outputs may still be incorrect, overconfident, repetitive, or unsafe.

## License

Apache 2.0, following the base checkpoint. See `LICENSE` and the upstream
[Qwen3.6-35B-A3B model card](https://huggingface.co/Qwen/Qwen3.6-35B-A3B).