← back to catalog · registered 2026-10-13 00:59

sam33339999/HuiHui-Qwen3.8-27B-Coder390-abliterated-MTP-Exl3-3.5bpw

sam33339999 27B multimodal second-order
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/sam33339999%2FHuiHui-Qwen3.8-27B-Coder390-abliterated-MTP-Exl3-3.5bpw"
Response includes
  • classification unknown
  • files 13
  • author_summary 2 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
?
Primary method

Unclassified

No clear signals of an abliteration technique in this model.
Confidence
UNKNOWN
Why this label 1 signal
No classification signals present. This may not be an abliterated model at all - it could be a repackaging, a merge with unrelated goals, or unrelated content that mentions the term.
  • no classification signals present (no abliterated, uncensored, or known producer/method markers)
Refusal direction extraction

No specific extraction method could be identified for this model. The producer either did not document it or used a proprietary pipeline.

What is a refusal direction? →
Downloads · 30-day
0
Likes
0
Model age
today
created 2026-10-13

Genealogy 0 direct forks

Full fork graph →

This model's place in the market. Above: what it was derived from. Below: the tree of everything derived from it.

Metadata

License
apache-2.0
Languages
en zh
Tags
safetensors qwen3_5 exl3 qwen3.8 mtp speculative-decoding multimodal abliterated exllamav3 image-text-to-text conversational en

Related

Total size
14.3 GB
Files
13
Quantizations
1
Registered
2026-10-13 00:59
Last updated on HF
2026-10-13 01:28

Files by quantization

Auxiliary files 13 files 14.3 GB
model-00001-of-00002.safetensors 7.94 GB 6401eda0 download
model-00002-of-00002.safetensors 6.34 GB 4da353a6 download
tokenizer.json 19.1 MB 06b95093 download
quantization_config.json 614 KB 882bc6c3 download
model.safetensors.index.json 233 KB 356c26ec download
chat_template.jinja 8.74 KB c0c686f9 download
config.json 4.70 KB c9512103 download
README.md 3.04 KB 22397deb download
.gitattributes 1.53 KB 52373fe2 download
tokenizer_config.json 1.05 KB 0f19934e download
preprocessor_config.json 390 B 2ea84a43 download
video_preprocessor_config.json 385 B 3ba673a5 download
generation_config.json 214 B c53835dc download

README current version from Hugging Face


license: apache-2.0
base_model: huihui-ai/Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated
base_model_relation: quantized
pipeline_tag: image-text-to-text
language:

  • en
  • zh
    tags:
  • exl3
  • qwen3.8
  • qwen3_5
  • mtp
  • speculative-decoding
  • multimodal
  • abliterated
  • exllamav3

Qwen3.8-27B Coder390 消融版 — EXL3 3.5 bpw

huihui-ai/Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated 的 BF16 權重,用 exllamav3 1.4.2 轉成的 EXL3。消融方式見那張模型卡:只動了第 2 到 52 層。這份卡只說明這次量化。

這是一個可以單獨使用的檢查點。文字塔、視覺塔和 MTP 頭都在同一個目錄裡。沒有附 DFlash2。

這個 EXL3 沒有重跑評測。來源模型卡上的分數是消融前的 BF16,不是這份 3.5 bpw。

量化

來源目錄的 model-00001、model-00002,加上索引裡的 mtp-bf16.safetensors。2026-10-13 在 NVIDIA GB10 上轉完。推理目標是 RTX 3090(24 GB)。

部分 位元率
文字塔線性層 平均 3.5 bpw(校正 250 列 × 2048 token,codebook mul1,output scales 始終開啟)
lm_head 6
MTP 的 8 個線性層 4。文字塔量化完才壓,寫進同一份 safetensors
視覺塔、embedding、norm BF16

目錄大約 14.3 GiB。架構仍是 Qwen3_5ForConditionalGeneration:64 層混合注意力(每 4 層一次 full attention),hidden 5120,詞表 248320,上下文 262144。generation_config.json 裡是 temperature 1.0、top_p 0.95、top_k 20。

MTP 的 39 個張量和視覺塔的 333 個張量都在 model-00002-of-00002.safetensors。輸出目錄沒有獨立的 mtp-*.safetensors。

怎麼載入

MTP 頭已經在這個目錄裡。下面的路徑換成你放權重的目錄。

RTX 3090 的三個旗標。直接呼叫時加上 --max_model_len 262144,對外公布的上下文才是原生的 262144:

python -u tools/serve_openai.py \
  -m /path/to/qwen3.8-27b-coder390-effthink-abliterated-EXL3-3.5bpw \
  -dm mtp \
  -cs 270336 \
  -cq 3 \
  --max_model_len 262144 \
  --host 127.0.0.1 \
  --port 5000
  • -dm mtp:用這個檢查點裡的 MTP 頭做草稿。mtp 不是路徑,不會再載第二個模型。預設窗口是 4 個 token。
  • -cs 270336:KV cache 的 token 格數。262144 是模型上下文,多 8192 給推測解碼的驗證窗。必須是 256 的倍數。
  • -cq 3:K 和 V 都是 3 bit。不寫就是 fp16,262144 在 24 GB 上放不下。fp8 和 nvfp4 配上這個長度也分配不出來。

start.sh 的 .env。DRAFT=mtp、CACHE_QUANT=3 就是上面兩個旗標。CONTEXT_SIZE=262144 時,start.sh 會自己把 cache 設成 270336:

MODEL_DIR=/path/to/qwen3.8-27b-coder390-effthink-abliterated-EXL3-3.5bpw
DRAFT=mtp
CONTEXT_SIZE=262144
CACHE_QUANT=3

GPU_MEM_GB 留空時,24 GB 卡會用 22。

視覺塔的 BF16 權重在這個檢查點裡。這次轉換沒有跑過圖像輸入。

授權

Apache-2.0,沿用來源的消融模型。

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