← back to catalog · registered 2026-10-23 03:00

sam33339999/Sage-A1.5-Heretic-Abliterated-Uncensored-exl3-4.0bpw

sam33339999 MoE 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%2FSage-A1.5-Heretic-Abliterated-Uncensored-exl3-4.0bpw"
Response includes
  • classification m3
  • files 15
  • author_summary 5 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M3
Primary method

Layer-wise ablation

Applied on top of direct removal inherited from the base model.
Confidence
HIGH
Inherited from base model
Why this label 2 signals
Producer identity confirmed by naming conventions, tags or the model card. This label is very unlikely to change.
  • 'heretic' in model name (Heretic-produced)
  • Heretic uses layer-wise optimization (M3) with underlying direction removal (M1)
Refusal direction extracted via
Extraction technique

Difference-of-means

Confidence
HIGH
Why we say so
name contains 'heretic'; Heretic default extraction is difference-of-means (Arditi 2024)
Downloads · 30-day
0
Likes
0
Model age
today
created 2026-10-23

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
Tags
exllamav3 safetensors qwen3_5_moe exl3 quantized moe vision mtp heretic abliterated uncensored image-text-to-text

Related

Total size
18.4 GB
Files
15
Quantizations
1
Registered
2026-10-23 03:00
Last updated on HF
2026-10-23 02:20

Files by quantization

Auxiliary files 15 files 18.5 GB
model-00002-of-00003.safetensors 7.93 GB 1d885d7d download
model-00001-of-00003.safetensors 7.69 GB 5734082b download
model-00003-of-00003.safetensors 2.77 GB 22634b21 download
quantization_config.json 38.5 MB 47afbb75 download
tokenizer.json 19.1 MB 2f4dd754 download
model.safetensors.index.json 13.1 MB 320cf1d5 download
chat_template.jinja 8.81 KB 4ed4d362 download
README.md 4.99 KB 5b94adf7 download
config.json 4.03 KB 093a9146 download
.gitattributes 1.66 KB f47dd0a1 download
tokenizer_config.json 1.43 KB e51a22c0 download
processor_config.json 1.27 KB 7ad6acdf download
preprocessor_config.json 390 B 2ea84a43 download
video_preprocessor_config.json 385 B 3ba673a5 download
generation_config.json 142 B 915b9e93 download

README current version from Hugging Face


library_name: exllamav3
license: apache-2.0
license_link: https://huggingface.co/InternScience/Sage-A1.5
base_model: sam33339999/Sage-A1.5-Heretic-Abliterated-Uncensored
base_model_relation: quantized
pipeline_tag: image-text-to-text
tags:

  • exl3
  • quantized
  • moe
  • vision
  • mtp
  • qwen3_5_moe
  • heretic
  • abliterated
  • uncensored

Sage-A1.5 Heretic Abliterated — EXL3 4.0 bpw

這個目錄是未審查 BF16 的 EXL3 量化,用 exllamav3 1.6.4 的 convert.py 產生。語言模型約 4 bpw,lm_head 6 bit,MTP 4 bpw,視覺塔保持 bf16。

模型關係:

InternScience/Sage-A1.5
  └─ sam33339999/Sage-A1.5-Heretic-Abliterated-Uncensored
       └─ 這個模型(EXL3 4.0 bpw)
階段 倉庫 是什麼
原始模型 InternScience/Sage-A1.5 原版 BF16
未審查 BF16 sam33339999/Sage-A1.5-Heretic-Abliterated-Uncensored 以原版做 Heretic 消融後的 BF16,已推上 Hub。這個 EXL3 的直接上游
這個模型 本目錄 上一行的 EXL3 4.0 bpw

base_model 只指向未審查 BF16。該倉庫自己的 base_model 已指向 InternScience/Sage-A1.5,Hub 的模型樹會顯示成「原始模型 → 未審查 BF16 → 這個量化」。

這不是 BF16 權重。原版或未審查 BF16 模型卡裡的 vLLM / SGLang 啟動指令不能拿來載入這個目錄。

量化設定

config.json 的 quantization_config:

欄位 值
quant_method exl3
version 1.6.4
bits 4.0(decoder;各層實測約 4.05 bpw)
head_bits 6
mtp_bits 4
vision_bits 無。視覺塔未量化,權重為 bf16
codebook mul1
out_scales auto
calibration 250 rows × 2048 cols,內建語料

沒有使用 --hq、per-tensor recipe、embedding 量化或 n-gram 表。

轉換指令:

python convert.py \
  -i /content/sage-a1.5-stage \
  -o /content/Sage-A1.5-Heretic-Abliterated-Uncensored-exl3-4.0bpw \
  -w /mnt/local-scratch/exl3-work-sage-a1.5 \
  -b 4 \
  -vb 16 \
  -mb 4 \
  -d 0

--vision_bits 16 是必要的。Qwen3VLVisionModel 的預設是 6 bpw,省略 -vb 16 會把視覺塔量化。-vb 16 時視覺權重經 get_additional_compiled_tensors 以 bf16 原樣寫入,quantization_config 不會出現 vision_bits。

權重來源

Decoder、視覺塔、tokenizer 與 chat_template.jinja 都來自未審查 BF16 sam33339999/Sage-A1.5-Heretic-Abliterated-Uncensored,不是直接從原版 InternScience/Sage-A1.5 量化。model.visual.* 保持 bf16,與該未審查 checkpoint 逐 bit 相同(已核對 model.visual.blocks.0.attn.proj.weight 與 model.visual.patch_embed.proj.weight)。

未審查 BF16 的分片沒有 mtp.*(text_config.mtp_num_hidden_layers 是 1,index 裡沒有這些鍵)。量化前只從原版 InternScience/Sage-A1.5 補了 model-mtp.safetensors。架構對過才使用這些張量,沒有改寫或另造權重。輸出含 mtp.fc.trellis 與 mtp.layers.0 的 expert trellis,沒有形狀 [2048, 4096] 的 bf16 mtp.fc.weight。MTP 依轉換器預設不做 calibration。

preprocessor_config.json、processor_config.json、video_preprocessor_config.json 未審查 BF16 沒有附上,同樣從原版補上,讓 Qwen3_5MoeForConditionalGeneration 能註冊 vision。影像前處理是 Qwen2VLImageProcessorFast,patch 16、temporal 2、merge 2。

目錄約 19GB,三個 safetensors shard。

使用

需已安裝能載入這個架構的 exllamav3(本次為 1.6.4)。文字生成:

python examples/chat.py \
  -m /content/Sage-A1.5-Heretic-Abliterated-Uncensored-exl3-4.0bpw \
  -mode qwen35

開啟同一目錄裡的 MTP draft:

python examples/chat.py \
  -m /content/Sage-A1.5-Heretic-Abliterated-Uncensored-exl3-4.0bpw \
  -mtp \
  -mode qwen35

-mode qwen35 是 examples/chat_templates.py 裡的 Qwen3.5 ChatML。模型自己的 chat_template.jinja 已原樣複製,程式載入時由 Tokenizer 使用。

程式載入:

from exllamav3 import Cache, Config, Generator, Model, Tokenizer

path = "/content/Sage-A1.5-Heretic-Abliterated-Uncensored-exl3-4.0bpw"
config = Config.from_directory(path)          # components: text, vision, mtp
model = Model.from_config(config)
tokenizer = Tokenizer.from_config(config)
cache = Cache(model, max_num_tokens=2048)     # 要在 load() 之前接上
model.load()
generator = Generator(model, cache, tokenizer)
print(generator.generate(prompt="The capital of France is", max_new_tokens=32, completion_only=True))

同一路徑連續兩次這樣載入都成功,短提示的補完非空。Cache 必須在 model.load() 之前建立,loader 才會一併配置 KV / recurrent state。

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