DeepSeek-V4.1-Flash Uncensored FP8 Released
At a Glance
| Item | Value |
|---|---|
| Repository | dealignai/DeepSeek-V4.1-Flash-UNCENSORED-FP8 |
| Published | 2026-09-11 |
| License | mit |
| Formats | safetensors |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code at collection time. Dates are JST.
Overview
dealignai has published “dealignai/DeepSeek-V4.1-Flash-UNCENSORED-FP8" on Hugging Face, which features permanent weight-level abliteration based on the multimodal model “DeepSeek-V4.1-Flash" with a 552B backbone parameter count. The information covered in this article is unconfirmed as official verification is not yet available, and future validation and trends require caution. A notable feature is that it can be loaded as a standard checkpoint without custom model.py or runtime hooks, surgically removing safety guardrails while maintaining up to 1M token context, vision capabilities, and multi-turn coherence.
Specifications
- License: mit
- Parameters: 552B backbone (active parameters 8B during prefill / 16B during decode)
- Context Length: Up to 1M tokens
- Architecture: Causal Encoder-Decoder (20+20 layers), MoE (384 routed experts top-6 + 1 shared), Hyper-Connections, CSA2 sparse attention, Engram n-gram memory, DSpark speculative draft
- Quantization: FP8 (
e4m3fn) weights with E8M0 block-scale [32, 32], FP4 routed experts
Performance
The performance of the original model (deepseek-ai/DeepSeek-V4.1-Flash) stated in the model card and the comparison table with frontier models are as follows.
→ Scroll horizontally to see all columns
| Benchmark (Metric) | # Shots | DeepSeek-V4-Flash-Base | DeepSeek-V4-Pro-Base | DeepSeek-V4.1-Flash-Base |
|---|---|---|---|---|
| Architecture | — | MoE | MoE | MoE |
| # Backbone Params | — | 284B | 1.6T | 552B |
| # Activated Params | — | 13B | 49B | 8B / 16B |
| World Knowledge | ||||
| AGIEval (EM) | 3–5-shot | 83.9 | 84.4 | 83.4 |
| MMLU-Pro (EM) | 5-shot | 68.3 | 73.5 | 74.1 |
| C-Eval (EM) | 5-shot | 92.1 | 93.1 | 92.1 |
| MultiLoKo (LLM-Judge) | 5-shot | 42.6 | 50.9 | 45.5 |
| SimpleQA-Verified (EM) | 25-shot | 30.1 | 55.2 | 42.3 |
| SuperGPQA (EM) | 5-shot | 46.5 | 53.9 | 53.1 |
| Language & Reasoning | ||||
| BBH (EM) | 3-shot | 86.9 | 87.5 | 86.1 |
| BBEH (EM) | 1-shot | 25.4 | 29.8 | 27.2 |
| DROP (F1) | 1-shot | 88.6 | 88.7 | 87.9 |
| HellaSwag (EM) | 0-shot | 85.7 | 88.0 | 87.2 |
| Code & Math | ||||
| BigCodeBench (Pass@1) | 3-shot | 56.8 | 59.2 | 60.6 |
| HumanEval (Pass@1) | 0-shot | 69.5 | 76.8 | 79.4 |
| GSM8K (EM) | 8-shot | 90.8 | 92.6 | 93.0 |
| MATH (EM) | 4-shot | 57.4 | 64.5 | 61.1 |
| MGSM (EM) | 8-shot | 85.7 | 84.4 | 80.2 |
| Long Context | ||||
| LongBench-V2 (EM) | 1-shot | 44.7 | 51.5 | 45.2 |
| Multimodal | ||||
| MMMU-Pro (EM) | 4-shot | — | — | 56.5 |
| CVBench (EM) | 4-shot | — | — | 77.9 |
| DocVQA (LLM-Judge) | 4-shot | — | — | 95.6 |
| RefCOCO-avg (Acc@0.5) | 0-shot | — | — | 86.0 |
Next, an excerpt of the comparison table with frontier models (maximum inference effort setting) is shown. Here, representative models and the original model group are contrasted.
→ Scroll horizontally to see all columns
| Benchmark (Metric) | Opus-5.0 | GPT-5.6 Sol | K3 | GLM-5.3 | DS-V4-Pro | DS-V4-Flash | DS-V4.1-Flash |
|---|---|---|---|---|---|---|---|
| Reasoning | |||||||
| GPQA Diamond (Pass@1) | 93.4 | 94.1 | 92.9 | 88.1 | 92.4 | 89.9 | 90.9 |
| HLE (Pass@1) | 56.3 | 44.5 | 43.5 | 42.0† | 42.7† | 37.8† | 36.8 (39.1†) |
| Codeforces (Rating) | — | — | — | — | 3348 | 3289 | 3471 |
| MathArena Apex (Pass@1) | — | — | 65.6 | — | 65.3 | 58.6 | 65.6 |
| Agentic | |||||||
| Terminal-Bench 2.1 (Pass@1) | 89.1 | 88.8 | 88.3 | 88.2 | 87.9 | 82.7 | 90.6 |
| Terminal-Bench 3.0 (Pass@1) | 43.3 | 34.4 | 17.7 | 28.3 | 11.8 | 7.6 | 30.0 |
| Terminal-Bench 4.0 (Pass@1) | 51.8 | 39.9 | 12.6 | 37.9 | 12.4 | 7.0 | 31.2 |
| DeepSWE v1.1 (Resolved) | 74.0 | 73.0 | 67.5 | 66.9 | 62.7 | 54.4 | 74.2 |
In measurements by the original model publisher, it records high scores such as 79.4 in HumanEval (Pass@1) and 60.6 in BigCodeBench (Pass@1) for coding, and demonstrates excellent performance in coding and math tasks with 93.0 in GSM8K (EM) for mathematics. It also reaches top levels, such as 3471 in competitive programming Codeforces (Rating) and 90.6 in agentic task Terminal-Bench 2.1. On the other hand, while it falls short of other models in HLE (Pass@1) which collects extremely difficult questions, publisher data indicates that this uncensored version disables guardrails while maintaining most of the high reasoning and coding capabilities possessed by the original model.
Strengths and Use Cases
Based on the specifications of the original model, its features include long-context processing up to 1M tokens, text and image multimodal inputs, tool calling, and advanced reasoning (continuously adjustable inference effort). This model is built with permanent weight-level abliteration, intended for use cases that generate substantive responses even to prompts categorized as harmful, such as chemistry and biology, cybercrime, weapons, self-harm, and harassment.
Differences from Similar Models
The base model was introduced in a previous article DeepSeek-V4.1-Flash Released, a 552B MoE Multimodal Model. The model covered in this article is an uncensored (abliterated) version that surgically and permanently removes the safety guardrails of that DeepSeek-V4.1-Flash, differing in that modifications have been applied to remove refusal behavior while retaining core components such as the base architecture, multimodality, 1M context length, and routed experts.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 763.2B parameters
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| More than 1083GB of VRAM (multi-GPU or CPU offload required) | BF16 | 902.8GB | 1083.3GB |
Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-09-18): llama.cpp: not registered, vLLM: registered, MLX (mlx-lm): not registered. “Not registered" means the name is absent from that registry today, not that the model cannot run.
Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
How to Get It
- Distribution Format: safetensors / FP8
- Supported Engines: SGLang (dsv4.1 preview branch) and vLLM (merge support in progress)
- Download Command (Hugging Face CLI, etc.):
huggingface-cli download dealignai/DeepSeek-V4.1-Flash-UNCENSORED-FP8