# DeepSeek-V4.1-Flash-Abliterated by Alex: Open-Weight Model
Source: https://savrn.com/models/deepseek-v4-1-flash-abliterated
Markdown alternate of the page above; the site index is https://savrn.com/llms.txt

---

## Runs On

What it takes to serve DeepSeek-V4.1-Flash-Abliterated (756.4B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.

| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
| --- | --- | --- | --- | --- | --- |
| 16-bit | 1512.8 GB | 1815.4 GB | 8x [MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) (256 GB) Vultr | $16.00 | [7x MI355X](https://savrn.com/ai-index/pricing/gpus/mi355x) $18.13 · [7x B300](https://savrn.com/ai-index/pricing/gpus/b300) $46.20 |
| 8-bit | 756.4 GB | 907.7 GB | 4x [MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) (256 GB) Vultr | $8.00 | [5x MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) $9.25 · [4x MI355X](https://savrn.com/ai-index/pricing/gpus/mi355x) $10.36 |
| 4-bit | 378.2 GB | 453.8 GB | 2x [MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) (256 GB) Vultr | $4.00 | [2x MI355X](https://savrn.com/ai-index/pricing/gpus/mi355x) $5.18 · [3x MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) $5.55 |

Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the [SAVRN Index](https://savrn.com/ai-index/pricing/gpus), read Oct 7, 2026.

[DeepSeek-V4.1-Flash-Abliterated on every accelerator the SAVRN Index prices, at every precision](https://savrn.com/models/deepseek-v4-1-flash-abliterated/gpus)

## Model Card

By Alex, published under mit, revision 48084075ae19.

[deepseek-ai/DeepSeek-V4.1-Flash](https://savrn.com/models/deepseek-v4-1-flash) with weight-level abliteration of the refusal direction.

### What was changed

The refusal direction was computed from 79 harmful vs. 79 benign instruction prompts (per-layer mean-difference of the collapsed residual stream, captured with the official reference implementation, tensor-parallel 4). Exactly 80 tensors were orthogonalized — for each of the 40 backbone layers:

- layers.N.attn.wo_b.weight — attention output projection (writes into the residual stream)
- layers.N.ffn.shared_experts.w2.weight — shared-expert down projection

Each weight W was edited as W ← W − r̂ (r̂ᵀ W) with r̂ the unit refusal direction of that layer, removing the model's ability to write the refusal direction into the residual stream. Weights were dequantized from FP8 [32×32] blocks (UE8M0 scales), edited in fp32, and requantized to the identical format.

Everything else is byte-identical to the base model: routed experts (FP4), Engram memory tables, CSA2 attention, router gates, norms, embeddings, the vision tower, and the DSpark draft head.

### Usage

[Read the full model card (266 words)](https://savrn.com/models/deepseek-v4-1-flash-abliterated/card)

## Configuration

Architecture

DeepseekV41ForCausalLM

Context length (tokens)

1,048,576

Layers

40

Hidden size

5,120

Attention heads

64

Key/value heads

1

Head dimension

512

Vocabulary size

129,280

Routed experts

384

Experts active per token

6

Sliding window (tokens)

128

RoPE base

10,000

Model type

deepseek_v41

Quantization

fp8

## Identity and Version

Repository

securepeak/DeepSeek-V4.1-Flash-Abliterated

Publisher

Alex

Task

Image and text to text

Modality

Image and text

Library

transformers

Parameters

756.4B parameters

Languages

moe

Revision

48084075ae1929403d7eadde4488427e665c4164

First published

2026-10-01

Last updated

2026-10-03

## Files and Weights

56 files, 765.6 GB in total. The weights are 48 files totalling 765.6 GB in safetensors.

Weights48 files · 765.6 GB

Configuration2 files · 7.5 MB

Tokenizer2 files · 6.4 MB

Documentation2 files · 3.4 KB

Other1 file · 1.8 MB

Repository1 file · 1.7 KB

Every file

| File | Type | Size | SHA-256 |
| --- | --- | --- | --- |
| model-00001-of-00048.safetensors | Weights | 970.5 MB | 00b8a36fc5ac |
| model-00002-of-00048.safetensors | Weights | 1.3 GB | b79f9953e915 |
| model-00003-of-00048.safetensors | Weights | 13.8 GB | a67cdc5b9d59 |
| model-00004-of-00048.safetensors | Weights | 13.8 GB | e9cc163fa9d8 |
| model-00005-of-00048.safetensors | Weights | 13.8 GB | 9287bb6da2c6 |
| model-00006-of-00048.safetensors | Weights | 13.8 GB | 2b58cd666c53 |
| model-00007-of-00048.safetensors | Weights | 13.8 GB | 73dd186e4e59 |
| model-00008-of-00048.safetensors | Weights | 13.8 GB | 57270d1e9c1b |
| model-00009-of-00048.safetensors | Weights | 13.8 GB | b82f80dbfd17 |
| model-00010-of-00048.safetensors | Weights | 13.8 GB | cf95fad8bfee |
| model-00011-of-00048.safetensors | Weights | 13.8 GB | 8ed025e47dd8 |
| model-00012-of-00048.safetensors | Weights | 13.8 GB | 197a47d7f2d9 |
| model-00013-of-00048.safetensors | Weights | 13.8 GB | 449d96b454b3 |
| model-00014-of-00048.safetensors | Weights | 13.8 GB | befa7bcaeec1 |
| model-00015-of-00048.safetensors | Weights | 13.8 GB | 869e9277150d |
| model-00016-of-00048.safetensors | Weights | 13.8 GB | 4e67ccfc2881 |
| model-00017-of-00048.safetensors | Weights | 13.8 GB | ef6139ed6be7 |
| model-00018-of-00048.safetensors | Weights | 13.8 GB | e10c4277af3b |
| model-00019-of-00048.safetensors | Weights | 13.8 GB | c378052edd10 |
| model-00020-of-00048.safetensors | Weights | 13.8 GB | 29fae17e0beb |
| model-00021-of-00048.safetensors | Weights | 13.8 GB | 587bc1d5f41d |
| model-00022-of-00048.safetensors | Weights | 13.8 GB | 37d931d189aa |
| model-00023-of-00048.safetensors | Weights | 13.8 GB | a8ecd6fba3d4 |
| model-00024-of-00048.safetensors | Weights | 13.8 GB | 9fd629b3a298 |
| model-00025-of-00048.safetensors | Weights | 13.8 GB | 4cac8388ff57 |
| model-00026-of-00048.safetensors | Weights | 13.8 GB | 4845ba6fc503 |
| model-00027-of-00048.safetensors | Weights | 13.8 GB | 605cb4687d25 |
| model-00028-of-00048.safetensors | Weights | 13.8 GB | 0c4b98185b0d |
| model-00029-of-00048.safetensors | Weights | 13.8 GB | 1b54898bf582 |
| model-00030-of-00048.safetensors | Weights | 13.8 GB | 1339a6d48baf |
| model-00031-of-00048.safetensors | Weights | 13.8 GB | add6182d51cd |
| model-00032-of-00048.safetensors | Weights | 13.8 GB | 542e1b7481e5 |
| model-00033-of-00048.safetensors | Weights | 13.8 GB | d244f107533d |
| model-00034-of-00048.safetensors | Weights | 13.8 GB | f461b956526c |
| model-00035-of-00048.safetensors | Weights | 13.8 GB | 9f9984863490 |
| model-00036-of-00048.safetensors | Weights | 13.8 GB | af1fd9e3b4a7 |
| model-00037-of-00048.safetensors | Weights | 13.8 GB | 45dd5ce83d13 |
| model-00038-of-00048.safetensors | Weights | 13.8 GB | f3b2d047dc2b |
| model-00039-of-00048.safetensors | Weights | 13.8 GB | 947a0fc9e073 |
| model-00040-of-00048.safetensors | Weights | 13.8 GB | 883aeb14ddfb |
| model-00041-of-00048.safetensors | Weights | 13.8 GB | 3e9499d68002 |
| model-00042-of-00048.safetensors | Weights | 13.8 GB | 07dee6550768 |
| model-00043-of-00048.safetensors | Weights | 1.3 GB | 616f7c844e5c |
| model-00044-of-00048.safetensors | Weights | 2.7 GB | f1b4c2780d69 |
| model-00045-of-00048.safetensors | Weights | 2.6 GB | 1adc78b4f76d |
| model-00046-of-00048.safetensors | Weights | 2.7 GB | e2f2e27ce6e0 |
| model-00047-of-00048.safetensors | Weights | 101.5 GB | 6d452740c150 |
| model-00048-of-00048.safetensors | Weights | 101.5 GB | 6b275888d26a |
| config.json | Configuration | 3.3 KB | — |
| model.safetensors.index.json | Configuration | 7.5 MB | — |
| LICENSE | Documentation | 1.1 KB | — |
| README.md | Documentation | 2.3 KB | — |
| DeepSeek_V41_Tech_Report.pdf | Other | 1.8 MB | ba68e2e40408 |
| .gitattributes | Repository | 1.7 KB | — |
| tokenizer.json | Tokenizer | 6.4 MB | — |
| tokenizer_config.json | Tokenizer | 801 B | — |

## License and Download

License

mit

Access

Open weights, no gate

Download size

765.6 GB

[Download from Alex](https://huggingface.co/securepeak/DeepSeek-V4.1-Flash-Abliterated)

Released by Alex through its official repository on Hugging Face. [Read the license](https://opensource.org/license/mit).

## Built From

- Derived from [deepseek-ai/DeepSeek-V4.1-Flash](https://savrn.com/models/deepseek-v4-1-flash)
- Quantized from [deepseek-ai/DeepSeek-V4.1-Flash](https://savrn.com/models/deepseek-v4-1-flash)

## Memory Requirements

| Precision | Weights in memory |
| --- | --- |
| As published | 765.6 GB |
| 16-bit | 1512.8 GB |
| 8-bit | 756.4 GB |
| 4-bit | 378.2 GB |

Weights only, from the published parameter count; the key-value cache and runtime add to this.

## Questions About DeepSeek-V4.1-Flash-Abliterated

### How much GPU memory does DeepSeek-V4.1-Flash-Abliterated need?

About 1815.4 GB at 16-bit and 453.8 GB at 4-bit: the weights (756.4B parameters) plus a working margin. A long context needs more.

### What is the cheapest GPU to run DeepSeek-V4.1-Flash-Abliterated on?

At 16-bit, 8x MI325X from $16.00 an hour; at 4-bit, 2x MI325X from $4.00 an hour, at the lowest on-demand prices the SAVRN Index lists.

### Can I use DeepSeek-V4.1-Flash-Abliterated commercially?

Yes. DeepSeek-V4.1-Flash-Abliterated is released under MIT License. The MIT License is a short permissive license. It permits commercial use, modification and redistribution, provided the copyright notice and permission notice are included.

### What is DeepSeek-V4.1-Flash-Abliterated's context length?

1,048,576 tokens, from the maximum position embeddings in its published configuration.

## Similar Models

Model · Image and text to text

### [DeepSeek-V4.1-Flash](https://savrn.com/models/deepseek-v4-1-flash)

[DeepSeek](https://savrn.com/model-publishers/deepseek-ai)

We introduce DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts (MoE) model with 552B backbone parameters and support for contexts of up to one million tokens. The model natively processes images and text, and generates text autoregressively. Architecture. DeepSeek-V4.1-Flash adopts a Causal Encoder-Decoder (CED) architecture: a 40-layer Transformer organized as a 20-layer causal encoder followed by a 20-layer decoder. With CED, the decoder's global KV cache is projected from the final encoder hidden states rather than derived from each decoder layer's own hidden states. This allows the model to activate only 8B parameters per token during prefill and 16B during decode, substantially…

Open weights mit 763.2B parameters 1,048,576 tokens transformers

[View model](https://savrn.com/models/deepseek-v4-1-flash)

Model · Image and text to text

### [DeepSeek-V4.1-Flash-UNCENSORED-FP8](https://savrn.com/models/deepseek-v4-1-flash-uncensored-fp8)

[SAIFI INDUSTRIES](https://savrn.com/model-publishers/saifiindustries)

with permanent weight-level abliteration — the safety guardrails have been surgically removed while preserving MMLU capability, vision, reasoning, MTP (DSpark), and multi-turn coherence. Proprietary weight-level abliteration developed by the dealignai research team. No custom model.py, no runtime hooks, no steering vectors — it's a standard checkpoint that loads exactly like the base model. The refusal circuitry is surgically removed while every capability-critical component (routed experts, Engram memory, CSA2 sparse attention, DSpark draft head, vision tower, router gates, norms, embeddings) is preserved byte-identical to the base. Every response 4-tier graded (HARDREF / SOFTRED / HEDGE /…

Open weights mit 763.2B parameters 1,048,576 tokens transformers

[View model](https://savrn.com/models/deepseek-v4-1-flash-uncensored-fp8)

Model · Image and text to text

### [s](https://savrn.com/models/s)

[Lautaro Rodriguez](https://savrn.com/model-publishers/user2simion2)

We introduce DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts (MoE) model with 552B backbone parameters and support for contexts of up to one million tokens. The model natively processes images and text, and generates text autoregressively. Architecture. DeepSeek-V4.1-Flash adopts a Causal Encoder-Decoder (CED) architecture: a 40-layer Transformer organized as a 20-layer causal encoder followed by a 20-layer decoder. With CED, the decoder's global KV cache is projected from the final encoder hidden states rather than derived from each decoder layer's own hidden states. This allows the model to activate only 8B parameters per token during prefill and 16B during decode, substantially…

Access requested at publisher mit 763.2B parameters transformers

[View model](https://savrn.com/models/s)

Model · Image and text to text

### [Synin-V1.1-Flash](https://savrn.com/models/synin-v1-1-flash)

[Synin AI Lab](https://savrn.com/model-publishers/synin)

We introduce DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts (MoE) model with 552B backbone parameters and support for contexts of up to one million tokens. The model natively processes images and text, and generates text autoregressively. Architecture. DeepSeek-V4.1-Flash adopts a Causal Encoder-Decoder (CED) architecture: a 40-layer Transformer organized as a 20-layer causal encoder followed by a 20-layer decoder. With CED, the decoder's global KV cache is projected from the final encoder hidden states rather than derived from each decoder layer's own hidden states. This allows the model to activate only 8B parameters per token during prefill and 16B during decode, substantially…

Open weights mit 763.2B parameters 1,048,576 tokens transformers

[View model](https://savrn.com/models/synin-v1-1-flash)

Model · Image and text to text

### [Qwen3.5-397B-A17B-FP8](https://savrn.com/models/qwen3-5-397b-a17b-fp8)

[Qwen](https://savrn.com/model-publishers/qwen)

Over recent months, we have intensified our focus on developing foundation models that deliver exceptional utility and performance. Qwen3.5 represents a significant leap forward, integrating breakthroughs in multimodal learning, architectural efficiency, reinforcement learning scale, and global accessibility to empower developers and enterprises with unprecedented capability and efficiency. For more details, please refer to our blog post Qwen3.5. WMT24++: a harder subset of WMT24 after difficulty labeling and rebalancing; we report the averaged scores on 55 languages using XCOMET-XXL. Empty cells (--) indicate scores not yet available or not applicable. MathVision：our model’s score is…

Open weights apache-2.0 403.4B parameters 262,144 tokens transformers

[View model](https://savrn.com/models/qwen3-5-397b-a17b-fp8)

Model · Image and text to text

### [GLM-5.3-Flash](https://savrn.com/models/glm-5-3-flash)

[Z.ai](https://savrn.com/model-publishers/zai-org)

Join our WeChat or Discord community. Check out the GLM-5.3-Flash blog and GLM-5 Technical report. Use GLM-5.3-Flash API services on Z.ai API Platform. We introduce GLM-5.3-Flash, the first natively multimodal model in the GLM-5 series. With 320B total parameters and just 18B active parameters, it outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price, while approaching Claude Opus 4.8 on coding and agentic benchmarks. GLM-5.3-Flash starts from a newly trained base model, with its architecture and training recipe redesigned around capability and efficiency. For the first time in the GLM series, we introduce a hybrid architecture combining sparse and linear…

Open weights mit 321.3B parameters 1,048,576 tokens transformers

[View model](https://savrn.com/models/glm-5-3-flash)

## Alex

[All models and datasets](https://savrn.com/model-publishers/securepeak)

## Versions

- [48084075ae19](https://savrn.com/models/deepseek-v4-1-flash-abliterated/versions/48084075ae19) · current 2026-10-03

## Explore More

- [All image and text to text models](https://savrn.com/models/tasks/image-and-text-to-text)
- [All models under mit](https://savrn.com/models/licenses/mit)
- [Model comparisons](https://savrn.com/models/comparisons)
- [The model directory](https://savrn.com/models)
- [Open model prices by host](https://savrn.com/ai-index/pricing/open-models)

## Source

- Repository metadata, read 2026-10-03.
- [Hugging Face record](https://huggingface.co/securepeak/DeepSeek-V4.1-Flash-Abliterated)
- [How the hub is built](https://savrn.com/model-hub/methodology)