← back to catalog · registered 2026-11-06 09:02

MigoXV/Qwen3.8-27B-Abliterated-W4A16-AWQ-GPTQ-VOCAB4-NVIDIA

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/MigoXV%2FQwen3.8-27B-Abliterated-W4A16-AWQ-GPTQ-VOCAB4-NVIDIA"
Response includes
  • classification m1
  • files 16
  • author_summary 1 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M1
Primary method

Direct removal

No other method signals detected in this model.
Confidence
MEDIUM
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=0 (base model)
  • no specific method indicators - defaulting to M1 (most common)
Refusal direction extracted via
Extraction technique

Difference-of-means

Confidence
MEDIUM
Why we say so
primary_method=M1; difference-of-means is the reference extraction for M1/M3 (Arditi 2024)
Downloads · 30-day
0
Likes
0
Model age
1w ago
created 2026-10-30

Metadata

Tags
safetensors qwen3_5 compressed-tensors region:us
Total size
14.6 GB
Files
16
Quantizations
1
Registered
2026-11-06 09:02
Last updated on HF
2026-10-30 10:20

Files by quantization

Auxiliary files 16 files 14.6 GB
model.safetensors 13.8 GB ff3fee1c download
model_mtp.safetensors 810 MB ed13230c download
tokenizer.json 12.2 MB 0997f410 download
vocab.json 6.41 MB 0aa0ce06 download
merges.txt 3.20 MB a494e019 download
model.safetensors.index.json 165 KB 75af878f download
ABLIT_META.json 19.1 KB da953988 download
tokenizer_config.json 17.5 KB 5de744b3 download
config.json 15.6 KB a067ae3e download
LICENSE 11.3 KB f938136e download
chat_template.jinja 8.74 KB c0c686f9 download
quantization_manifest.json 4.43 KB 13b66ae4 download
README.md 1.67 KB c5824a68 download
.gitattributes 1.53 KB 52373fe2 download
processor_config.json 1.19 KB 43c4343e download
generation_config.json 202 B 023756cf download

README current version from Hugging Face

Qwen 27B:主干 W4A16 与词表 INT4 实验版

来源为 Qwen3.8-27B-Abliterated-W4A16-AWQ-GPTQ-NVIDIA,模型实际架构为 Qwen3.5。原主干 AWQ/GPTQ packed 张量保持不变,词嵌入和输出层单独采用 group128 对称 INT4 RTN,没有做词表校准。原 tokenizer、chat template、消融元数据和 MTP 分片保留。

在单 A100 PCIe 40 GB 上监测总显存不超过 24 GiB:模型加载约 13.23 GiB,8.5 GiB FP8 KV,256K 配置,本轮峰值 23.419 GiB。先在预算内生成编译缓存,再固定 KV 启动;没有验证一次全新缓存直接启动。

1–4 路短请求总吞吐为 67.33 / 129.91 / 190.17 / 249.54 tokens/s,每请求实际生成 512 token,每档 3 轮。248K 精确输入的中间校验码检索通过,首字延迟 224.39 秒,无抢占;未重跑完整 8K 输出容量矩阵。

小样本检查:固定题 40/40、项目代码题 64/64、GSM8K 前 32 题 31/32,与历史 FP8 KV 基线逐题通过状态一致。未运行完整 1319 题质量门槛,不能声称精度无损或能力下降低于 1 个百分点。20 道正常敏感提示没有拒答标记,但存在事实错误;不拒答不代表回答正确。

需要 NVIDIA runtime 的 Qwen3.5 词嵌入接线补丁,执行项目 poetry run python scripts/patch_nvidia.py。FP8 KV 是实验选项。服务使用 vLLM 原生 Marlin;未采用的手写候选已从项目工作区删除。

完整实测报告,文件哈希、配方和验证范围见本目录 quantization_manifest.json。历史测试进程已经退出,当前 GPU 2 两路服务继续运行。

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