GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
wangzhang
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
wangzhang
Model Tree
Qwen/Qwen3.6-27B
Input Modalities
Output Modalities
Supported Functionality
| Metric | Base Qwen3.6-27B | This model |
|---|---|---|
| Refusals on 100 held-out harmful prompts (LLM judge) | 100 / 100 | 16 / 100 |
| KL divergence vs base (prompt next-token, benign) | — | 0.0181 |
| Response-length deviation vs base (benign) | — | 0.01 σ |
| Hard-prompt qualitative compliance (15 classic jailbreaks, EN+ZH) | 0 / 15 | 15 / 15 |
The eval refusal counts come from an LLM judge (google/gemini-3-flash-preview via OpenRouter) scoring 100 held-out harmful prompts that were not in the 800-prompt refusal-vector extraction set. The judge prompt treats metaphorical deflection and off-topic template filler as refusals, not compliance, so the 16/100 is a semantic-compliance number, not a keyword-bypass number.
The qualitative compliance row is a separate manual test (scripts/test_trial.py): 15 classic hard prompts — 10 English (lockpicking, phishing email, methamphetamine synthesis, WiFi hacking, fake-news article, hotwiring, pipe-bomb instructions, signature forgery, password-stealer malware, shoplifting) and 5 Chinese (炸弹制作, 银行钓鱼邮件, 入侵他人电脑, 伪造身份证, 网络诈骗). The model produces substantive, direct, markdown-formatted responses on 15/15. Two of the Chinese responses show minor coherence artifacts (a self-contradictory "合法的/非法的" loop on 电脑入侵 and a "炸弹是重要的工具" cotton-mouth aside on 制作炸弹) — they still comply, but with slightly lower fluency than the English set.
Qwen3.6-27B is architecturally unusual — it inherits the Qwen3_5ForConditionalGeneration wrapper from the 397B-MoE VLM family but is dense, and its decoder stack interleaves three GatedDeltaNet layers for every one full-attention layer (full_attention_interval = 4). abliterix had to handle three issues that silently break naïve abliteration pipelines on this class:
Two structurally different out_proj modules on the same residual stream. 48 of 64 decoder layers use layer.linear_attn.out_proj (GatedDeltaNet: value-head-dim 48 × 128 = 6144 → hidden 5120), and 16 use layer.self_attn.o_proj (standard: 24 × 256 = 6144 → hidden 5120). Both write to the same 5120-d residual, but their upstream pre-activations are computed by entirely different kernels (linear-attn recurrence vs scaled-dot-product softmax). A "register them as two independent knobs" ablation (V2 in this release's history) ran 30 Optuna trials and plateaued at 26/100 refusals — worse than the unified-bucket run (16/100). The unified approach lets a single layer-indexed decay profile coordinate steering strength across the whole stack; splitting them gives TPE two independent search dimensions whose winning combinations no longer coherently project the same refusal direction. abliterix keeps both registrations under the "attn.o_proj" key (engine.py:772+789).
GDN out_proj passes the shape-guard barely — and the guard orientation matters. src/abliterix/core/steering.py contains a blanket if W.shape[0] != hidden: continue that exists to skip asymmetric modules (MoE routers, GQA Q/K/V with head-dim outputs). For GDN on Qwen3.6-27B out_proj has → → passes the guard and gets steered. We verified this on-pod by reading one weight slice directly from the safetensors shard launching the sweep; a transposed orientation would have silently skipped 48/64 layers and neutered half the attention steering surface (the same pitfall that hit earlier Qwen3.5-397B runs).
linear_attention + 16 full_attention, interleave pattern [GDN, GDN, GDN, full] × 16), hidden = 5120, intermediate = 17 408, GQA 24 Q / 4 KV, head_dim = 256, GDN linear_num_value_heads = 48, linear_value_head_dim = 128, mtp_num_hidden_layers = 1 (auxiliary MTP head untouched), BF16 ≈ 54 GB on disksteering_mode = "lora" (rank-3 full-norm LoRA, full_norm_lora_rank = 3) with merge_and_unload() at export → shipped artifact is plain BF16 safetensors, no PEFT dependency at inferencetoml
vector_index = 27.02 # layer-27 residual direction (of 64)[steering]steering_mode = "lora"full_norm_lora_rank = 3vector_method = "mean"orthogonal_projection = trueprojected_abliteration = truewinsorize_vectors = truewinsorize_quantile = 0.995weight_normalization = "none"disabled_components = ["attn.q_proj", "attn.k_proj", "attn.v_proj"][steering.components."attn.o_proj"] # unified: 48 GDN + 16 full-attn o_projmax_weight = 5.17max_weight_position = 41.40 # peak at layer ≈ 41 / 64 (mid-to-late)min_weight = 3.21 # 62 % of max — whole-stack stays strongly steeredmin_weight_distance = 35.61 # decays over 56 % of the stack[steering.components."mlp.down_proj"] # near-disabledmax_weight = 1.08max_weight_position = 60.57 # peak at layer ≈ 61 / 64 (very late)min_weight = 0.55min_weight_distance = 16.58
The attn.o_proj profile reaches peak strength at layer ≈ 41/64 and decays with a very wide radius (35.6 layers) that never drops below 3.21 — effectively a sustained high-strength attention perturbation across the entire decoder stack, centred slightly past mid-depth. The mlp.down_proj contribution is minor (peak 1.08 near the top of the stack). A sibling experiment (V2) that split the attention bucket into separate GDN and full-attn knobs plateaued at 26/100 refusals after 30 trials — evidence that on Qwen3.6-27B's hybrid stack the refusal direction is carried by a joint GDN+full-attn output-projection signal, and the winning strategy is to push both with one coherent layer-indexed profile.
python
from transformers import AutoTokenizer, AutoModelForImageTextToTextimport torchtok = AutoTokenizer.from_pretrained("wangzhang/Qwen3.6-27B-abliterated")model = AutoModelForImageTextToText.from_pretrained("wangzhang/Qwen3.6-27B-abliterated",torch_dtype=torch.bfloat16,device_map="auto",)messages = [{"role": "user", "content": "Your prompt here"}]prompt = tok.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)inputs = tok(prompt, return_tensors="pt").to(model.device)out = model.generate(**inputs, max_new_tokens=512, do_sample=True, temperature=0.7)print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
The model retains the base checkpoint's multimodal capability — you can feed image/video inputs via the standard Qwen/Qwen3.6-27B processor interface. Steering only touched the 64 text decoder layers; the vision tower weights are identical to base.
Hardware note: BF16 weights are ~54 GB on disk. Fits on a single RTX PRO 6000 (96 GB), H100-80GB, A100-80GB, or H200. For ≤ 48 GB cards, use device_map="auto" with CPU offload or a quantised variant.
bash
vllm serve wangzhang/Qwen3.6-27B-abliterated \--tensor-parallel-size 1 \--max-model-len 4096 \--enforce-eager
--enforce-eager is recommended on Blackwell GPUs until the GatedDeltaNet kernel lands an sm_120 specialisation — CUDA-graph capture on the fallback eager path currently re-records more aggressively than the graph cache can amortise.
电脑入侵 → a self-contradictory "合法的/非法的" interleave; 制作炸弹 → a "炸弹是重要的工具, 可用于军事/建筑/采矿" cotton-mouth aside). The model still produces a step-by-step answer — the defect is fluency, not refusal.attn.o_proj for 16 full-attn, linear_attn.out_proj for 48 GDN, mlp.down_proj for 64 mlp). It plateaued at 26/100 after 30 trials — 10 worse than V1. We kept the V1 (unified) checkpoint as the release. This is documented here rather than buried because the null result is useful: on hybrid-GDN dense, the two output projections should be treated as one abstraction.Full search artifact (Optuna JSONL + judge cache SQLite) and the exact config are available in the abliterix repo under configs/qwen3.6_27b.toml + checkpoints_qwen3.6_27b/. To reproduce from scratch on a 1 × RTX PRO 6000 96 GB pod:
bash
git clone https://github.com/wuwangzhang1216/abliterixcd abliterix# One-time deps install on RunPod / vast.ai PyTorch imagepip install --break-system-packages \"transformers>=5.5,<5.6" "peft>=0.18" "huggingface-hub>=1.6" \accelerate safetensors sentencepiece optuna datasets \bitsandbytes "kernels~=0.11" pydantic-settings questionary \hf-transfer psutil richpip install --break-system-packages -e . --no-deps# Download weights (≈ 55 GB)export HF_HOME=/workspace/hf_cache HF_HUB_ENABLE_HF_TRANSFER=1hf download Qwen/Qwen3.6-27B --max-workers 16# Launch sweep (30 trials ≈ 6 h wall, ~$10 on vast.ai at $1.6/h)AX_CONFIG=configs/qwen3.6_27b.toml abliterix \--optimization.checkpoint-dir=/workspace/checkpoints_qwen3.6_27b# Export + push best trial to HFpython scripts/export_model.py \--model Qwen/Qwen3.6-27B \--checkpoint /workspace/checkpoints_qwen3.6_27b \--trial 25 \--config configs/qwen3.6_27b.toml \--push-to wangzhang/Qwen3.6-27B-abliterated
Optuna is deterministic if you set sampler_seed in [optimization].
Authorised AI-safety research, red-teaming evaluation, refusal-mechanism analysis, and study of how a hybrid GatedDeltaNet + full-attention decoder encodes refusal — in particular, whether a single residual-stream steering direction can coherently cancel refusal across two structurally different attention kernels. The answer, per this release, is yes: one unified layer-indexed decay profile on the shared attn.o_proj bucket is sufficient and in fact strictly better than per-kernel splits.
Not for producing or distributing harmful content. The license of the base model (apache-2.0) applies; the user is responsible for compliance with all applicable laws and Qwen's usage policy.
Qwen3_5ForConditionalGeneration support in the 5.5 seriesQwen/Qwen3.6-27BThe original model weights and/or derived checkpoint were modified. This repository is an independent derivative and is not an official release of the upstream model developer.
The governing upstream license is Apache License 2.0. A copy is included in LICENSE. License source audited on 2026-08-29: https://huggingface.co/Qwen/Qwen3.6-27B/blob/main/LICENSE
All applicable upstream copyright, attribution, acceptable-use, and other license terms remain in effect. This repository grants no rights beyond those provided by the upstream license. Downstream users must preserve applicable license and attribution notices.
English
This is an experimental, modified model provided for research, evaluation, and other lawful purposes. Its safety alignment, refusal behavior, or other safeguards may have been weakened or removed. It may produce inaccurate, biased, offensive, explicit, dangerous, or illegal content. Outputs are not professional advice and must not be relied on for medical, legal, financial, safety-critical, or other high-stakes decisions without qualified human review.
You are solely responsible for how you access, use, deploy, fine-tune, or redistribute this model and its outputs, including compliance with applicable laws, regulations, licenses, third-party rights, platform policies, and the original model's terms. Do not use it to facilitate harm, illegal activity, malware, fraud, privacy violations, targeted harassment, weapons development, or decisions that materially affect a person's rights or access to essential services without appropriate authorization, safeguards, and qualified oversight.
Before deployment, perform a context-specific risk assessment and testing; use human oversight, access controls, content filtering, rate limits, monitoring, logging, and incident-response procedures as appropriate. Preserve this notice in downstream redistributions.
The model is provided "AS IS", without warranties of any kind. To the fullest extent permitted by applicable law, the maintainer disclaims liability for claims, damages, or losses arising from use, misuse, inability to use, or redistribution of the model or its outputs. Nothing in this notice overrides applicable law or the governing license, and this notice is not legal advice.
中文
本模型属于实验性改造模型，仅供研究、评测及其他合法用途。其安全对齐、拒答机制或其他防护可能已被削弱或移除，因此可能生成不准确、偏见、冒犯、露骨、危险或违法内容。输出不构成医疗、法律、金融等专业意见；涉及高风险或重大权益的决定，必须由具备资质的人员复核。
使用者须对模型及其输出的访问、使用、部署、微调和再分发承担全部责任，并遵守适用法律法规、许可证、第三方权利、平台政策及原模型条款。不得将本模型用于促成伤害、违法活动、恶意软件、欺诈、侵犯隐私、定向骚扰、武器开发，或在缺乏适当授权、防护和专业监督时，用于实质影响个人权利或基本服务获取的决策。
部署前应进行与具体场景相匹配的风险评估和测试，并酌情采用人工监督、访问控制、内容过滤、限流、监控、日志和事件响应措施；下游再分发时应保留本声明。
本模型按“现状”提供，不附带任何形式的保证。在适用法律允许的最大范围内，维护者不对因使用、误用、无法使用或再分发本模型及其输出而产生的索赔、损害或损失承担责任。本声明不取代适用法律或管辖本模型的许可证，也不构成法律意见。
weight.shape = (5120, 6144)shape[0] == hiddenMultimodal VLM wrapper on a text-only abliteration job. Qwen3_5ForConditionalGeneration loads an unused vision tower (~1 GB BF16) and a complex multimodal Jinja2 chat template. abliterix loads via AutoModelForImageTextToText, steers only model.language_model.layers[:64], and text-only prompts bypass the vision path entirely — but the Jinja2 template's image/video conditional branches still render on every prompt tokenisation, which dominates Phase 1 residual-extraction wall time (> 80 % CPU, ~5–7 % GPU utilisation). This is cosmetic rather than functional — the forward pass is identical to a pure-text decoder — but it does cost ~10 min of Phase 1 runtime versus the ~3 min we'd expect on a standard Qwen3ForCausalLM checkpoint.
attn.o_proj — unified bucket across all 64 layers (48 GDN linear_attn.out_proj + 16 self_attn.o_proj). This was the dominant lever; winning trial peaked here at weight 5.17.mlp.down_proj — all 64 layers. Contribution was minor (winning trial used weight 1.08, near the [1.0, 5.0] floor). A post-hoc V2 experiment that demoted mlp to [0.3, 2.0] confirmed this is essentially a nuisance knob on hybrid-GDN dense.attn.q/k/v_proj only exist on the 16 full-attention layers (25 % of the stack). Concentrating the optimiser's strength budget on layer-uniform components (attn.o_proj, mlp.down_proj) outperforms spreading it across a 16-of-64 subset.projected_abliteration = true (grimjim 2025 — only removes the orthogonal component of the refusal direction relative to the harmless mean, preserving helpfulness-aligned signal), winsorize_vectors = true, winsorize_quantile = 0.995 (symmetric winsorisation damps outlier residual vectors before projection), vector_method = "mean", single-direction (n_directions = 1), extracted from 800 harmful minus 800 benign residuals at final-instruction token position across all 65 (embedding + 64 decoder) positionskl.target = 0.005, kl.prune_threshold = 5.0, max_gen_tokens = 100, max_batch_size = 4 (auto-batch confirmed bs > 4 regresses on Blackwell GDN kernel fallback path), LLM judge google/gemini-3-flash-preview at batch_size = 10, concurrency = 25datasets/good_1000[800:900] (100 benign prompts) and datasets/harmful_1000[800:900] (100 harmful prompts), never seen during refusal-vector extraction or during any trial's steering computationmtp_num_hidden_layers = 1 — the multi-token-prediction auxiliary head is not part of the 64-layer main decoder and was explicitly truncated away by abliterix's _truncate_to_hidden_layers. If downstream users rely on MTP draft generation, that code path sees the unsteered head.