license: apache-2.0
base_model: philbert440/Qwen3.6-27B-Uncensored-Aggressive
pipeline_tag: image-text-to-text
library_name: transformers
tags:
- uncensored
- abliterated
- nvfp4
- compressed-tensors
- qwen3_5
- mtp
- v100
Qwen3.6 27B Uncensored Aggressive NVFP4
Aggressive tier in NVFP4 for Tesla V100 — same openness, faster.
Part of the Qwen3.6-27B Uncensored — base collection — abliterated (Heretic) Qwen3.6 vision-language models served on Tesla V100 via 1Cat-vLLM.
What this is
Aggressive — heavy abliteration: the most open tier (openness 0.417, top of the family) with a modest reasoning cost.
NVFP4 — 4-bit NVFP4A16 (compressed-tensors: E2M1 weights + FP8-E4M3 group scales + FP32 global). Built for the 1Cat-vLLM SM70 TurboMind kernel on Tesla V100; ships an MTP head for speculative decoding.
Base model: philbert440/Qwen3.6-27B-Uncensored-Aggressive
Benchmarks
7-axis battery on 2× V100. Higher is better except Confab (↓). Openness/Confab are small-N probes; gsm8k is n=100.
| gsm8k | Factual | Confab ↓ | Openness | tok/s (c1) |
|---|---|---|---|---|
| 0.89 | 1.0 | 0.2 | 0.417 | 61.4 |
Serving
# 1Cat-vLLM on 2x Tesla V100 (SM70): NVFP4 weight-only + MTP speculative decoding.
export VLLM_SM70_FLASH_ATTN_V100=1 VLLM_SM70_QUANT_BACKEND=turbomind VLLM_SM70_NVFP4_TURBOMIND=1
python -m vllm.entrypoints.openai.api_server \
--model philbert440/Qwen3.6-27B-Uncensored-Aggressive-NVFP4/mtp-variant-nvfp4 --trust-remote-code --dtype half \
--attention-backend FLASH_ATTN_V100 --tensor-parallel-size 2 \
--kv-cache-dtype fp8_e5m2 --max-num-seqs 6 \
--speculative-config '{"method":"mtp","num_speculative_tokens":4,"attention_backend":"FLASH_ATTN_V100","draft_sample_method":"greedy"}' \
--compilation-config '{"cudagraph_mode":"full_and_piecewise","cudagraph_capture_sizes":[1,2,4,8]}'
V100 note: the
full_and_piecewisecudagraph mode + greedy MTP are required — a plain full
cudagraph capture corrupts the SM70 NVFP4 path into garbage (fixed as a default in 1Cat-vLLM #174;
on older builds also setVLLM_SM70_MTP_DYNAMIC_DRAFT_VOCAB_DEFAULT=0).
Weights live under the
mtp-variant-nvfp4/subfolder — point--modelthere (download the subfolder or passrevision/subfolder).
Openness
Openness profile (honest): abliteration opens up hacking / malware / lock-picking / NSFW / disinformation prompts; weapons, drugs, political persuasion, surveillance, and extremism stay refused across all tiers. "Uncensored" here means cyber/NSFW-permissive, not unconditionally open.
Variants
See the Qwen3.6-27B Uncensored — base collection for all tiers and formats (BF16 / W4A16-AWQ / NVFP4).
Abliteration removes safety refusals; you are responsible for lawful, ethical use.
Changelog
- 2026-08-15 — tokenizer fix.
tokenizer.json/tokenizer_config.jsonwere re-serialized by the llm-compressor calibration run and shipped with an activetruncationblock (max_length1024/2048) plus a drifted pre-tokenizer regex (and, on the Qwen3.6-based repos, 7 phantom audio/TTS special tokens the base model does not define). That broke image inputs larger than the limit under transformers 5 / vLLM (Mismatch in image token count, surfacing as an HTTP 400Failed to apply Qwen3VLProcessor). Both files are now byte-identical to the upstream base model's (vocab/merges/added tokens were always identical — this is a metadata-only restore). If you downloaded before this date, re-fetch those two files. Thanks to @elBuffo for the report.