Get the app
← back to catalog · registered 2026-11-09 22:03

momokokong/DeepSeek-V4.1-Flash-1017-EXL3-2.50bpw-Abliterated-TensorFold-TP2

momokokong Deepseek multimodal
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/momokokong%2FDeepSeek-V4.1-Flash-1017-EXL3-2.50bpw-Abliterated-TensorFold-TP2"
Response includes
  • classification m1
  • files 12
  • author_summary 1 models
  • readme_text full
Try next
GET/models?family=deepseek&method=m1
Sibling deepseek models · same method
 GET/methods/m1
What "m1" means · taxonomy
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)
How we label → About M1 →
Refusal direction extracted via
Extraction technique

Difference-of-means

FromArditi et al. (2024) ↗
Confidence
MEDIUM
Why we say so
primary_method=M1; difference-of-means is the reference extraction for M1/M3 (Arditi 2024)
What is a refusal direction? → Read the paper ↗
Downloads · 30-day
23
Likes
0
Model age
today
created 2026-11-09
Downloads over time
Now0→from0↑0%
00110 on Nov 90 on Nov 10Nov
Nov 9 → Nov 10 · 2 snapshots · spans 1 day

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.

deepseek-ai/DeepSeek-V4.1-Flash-1017 this lineage
momokokong/DeepSeek-V4.1-Flash-1017-EXL3-2.50bpw-Abliterated-TensorFold-TP2↓ 23

Metadata

License
mit
Tags
safetensors deepseek_v41 exl3 tensorfold dgx-spark gb10 abliterated uncensored deepseek-v4.1 multimodal image-text-to-text conversational

Related

Same base model
huihui-ai/Huihui-DeepSeek-V4.1-Flash-1017-abliterated-GGUF 66K AtomicChat/DeepSeek-V4.1-Flash-1017-Uncensored-DwarfStar-GGUF 47K IndrasMirror/DeepSeek-V4.1-Flash-1017-XQ-Uncensored 28K dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8 17K AtomicChat/DeepSeek-V4.1-Flash-1017-Uncensored-GGUF 13K

Control vectors

Vectors for this model →
msuiche/DeepSeek-V4.1-Flash-1017-abliterated-cyber-GLP-27-L11-37-a3.5 1
Total size
12.7 GB
Files
12
Quantizations
1
Registered
2026-11-09 22:03
Last updated on HF
2026-11-09 21:51
Source huggingface.co/momokokong/DeepSeek-V4.1-Flash-1017-EXL3-2.50bpw-Abliterated-TensorFold-TP2 ↗ Base model deepseek-ai/DeepSeek-V4.1-Flash-1017 ↗

Files by quantization

Auxiliary files 12 files 12.8 GB
model-00001-of-00015.safetensors 6.74 GB 47aead29 download
model-00002-of-00015.safetensors 6.00 GB 693de311 download
model.safetensors.index.json 14.1 MB 859525e6 download
tokenizer.json 6.07 MB 6a15814d download
engram_token_map.u32 505 KB c60a8632 download
config.json 4.34 KB a3975930 download
chat_template.jinja 3.92 KB 37c14134 download
README.md 3.91 KB 3766f940 download
.gitattributes 1.60 KB 3f4617d8 download
LICENSE 1.06 KB d62e3bef download
tokenizer_config.json 801 B f3dad388 download
generation_config.json 191 B 979fe07f download

README current version from Hugging Face

license: mit
base_model:

  • deepseek-ai/DeepSeek-V4.1-Flash-1017
  • vcruz305/DSV4.1-Flash-1017-DERISKED-EXL3-2.50bpw
  • bertholomus/DeepSeek-V4.1-Flash-1017-EXL3-2.5bpw
  • drowzeys/keys-DeepSeek-V4.1-Flash-1017-Abliterated
    tags:
  • exl3
  • tensorfold
  • dgx-spark
  • gb10
  • abliterated
  • uncensored
  • deepseek-v4.1
  • multimodal
    pipeline_tag: image-text-to-text

DeepSeek-V4.1-Flash-1017-EXL3-2.50bpw-Abliterated-TensorFold-TP2 (Delta Pack for 2× DGX Spark)

GitHub Recipe & Scripts

This repository hosts the 12.74 GiB Pre-built Delta Pack (model-00001-of-00015.safetensors + model-00002-of-00015.safetensors + model.safetensors.index.json) for running DeepSeek-V4.1-Flash-1017 (Abliterated / Uncensored) at 2.50 bpw EXL3 (TP=2) across 2× NVIDIA DGX Spark (GB10 128 GB) nodes using the native TensorFold engine.

For the full deployment recipe, automated zero-copy linker script (download_and_build.sh), systemd service, and watchdog scripts, visit the companion GitHub repository:
👉 https://github.com/momokokong/DeepSeek-V4.1-Flash-1017-EXL3-2.50bpw-Abliterated-TensorFold-TP2

What Is Included in This Delta Pack?

File Size Description
model-00001-of-00015.safetensors 6.74 GiB (7,237,136,672 B) 3,571 non-expert & abliterated tensors:
• Base 2.5bpw EXL3 non-expert tensors from bertholomus/DeepSeek-V4.1-Flash-1017-EXL3-2.5bpw
• Grafted 104 abliterated layers.10..35.attn.wo_b EXL3 tensors (651 MiB) from drowzeys/keys-DeepSeek-V4.1-Flash-1017-Abliterated
• TensorFold compatibility fixes: mtp.2.markov_head.{embed,head}.weight (native BF16/F16), layers.{1,14}.engram.{q,k}_weight widened to F32, and all 43 ffn.gate.bias_vl vision routing biases
model-00002-of-00015.safetensors 6.00 GiB (6,442,515,560 B) 8,519 pure routed-expert EXL3 tensors extracted from vcruz305/DSV4.1-Flash-1017-DERISKED-EXL3-2.50bpw shard 2 (with the 218 incompatible FP8 non-expert tensors stripped)
model.safetensors.index.json 14.7 MB Complete 15-shard index (192,499 keys, 167.99 GiB total) mapping shards 00001..00002 (from this repo) and shards 00003..00015 (155.25 GiB, hardlinked from vcruz305/DSV4.1-Flash-1017-DERISKED-EXL3-2.50bpw)
config.json & Tokenizers 6.9 MB TensorFold-compatible EXL3 configuration, engram_token_map.u32, chat_template.jinja, and tokenizer files

Why Use This 2.50bpw Hybrid Pack Over 2.9bpw?

  1. Saves ~14 GiB of Unified Memory Per Node:
    • Reduces the per-node TensorFold rank-cache from 98.85 GiB (2.9bpw) down to 84.72 GiB (89.92 GiB total resident footprint).
  2. ~16.5 – 17.5 GiB Free MemAvailable on Both Sparks:
    • Even at CONTEXT=250000, POOL=524288, and PARALLEL=4, both nodes maintain 16.5+ GiB of free RAM during 78K long-context stress tests—allowing you to co-locate Agent Gateways, local Embedding/Reranker servers, and ComfyUI without triggering memguard or OOM kills.
  3. Full Speed, Vision & Speculative Decoding:
    • Cold Boot Time: 33.7 s (81.2 GiB expert cache read in 16.2 s)
    • Single-Stream Decode: 64.25 tok/s median (tools/bench.py --suite decode), up to 90.51 tok/s on code generation
    • 4-Way Concurrent Decode: 124.62 tok/s (tools/bench.py --suite multi), up to 153.58 tok/s on 4-way structured generation
    • Cold Prefill (cached=0): 2,134.8 tok/s @ 8K | 2,184.4 tok/s @ 32K | 2,152.0 tok/s @ 78K
    • Needle in a Haystack (32K & 78K, depths 0.25, 0.50, 0.75): 6 / 6 (100% CORRECT)

README history 1 version

The author's README evolved over time. Click a version to see its content at that point.

  1. 2026-11-09Add Model Card and 2x DGX Spark benchmark results6c120503.9 KB
    Loading...
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.

Get the app Browse the catalog
Open in Abliteration