DeepSeek V4.1 Flash Uncensored

• 🔒 Community abliterated fork of DeepSeek V4.1 Flash; refusal behavior removed
• 🧠 Multimodal MoE, 552B backbone params, 8B active on input, 16B on output
• 📏 Served here at 512K context; upstream card cites up to 1M
• 👁️ DeepSeek-ViT vision tower and projector preserved byte-identical
• 🔧 Tool calling, reasoning-effort tiers, DSpark speculative decoding retained
• ⚡ Native FP8 weights with FP4 routed experts, drop-in checkpoint
• 📚 MIT-licensed base, enabling redistribution and modification
• 🎯 No content-filtering guardrails — unsuitable for consumer-facing defaults

DeepSeek

DeepSeek is a Chinese artificial intelligence company specializing in large language model development, founded in July 2023 by Liang Wenfeng. Based in Hangzhou, Zhejiang, the company is backed by High-Flyer, a prominent Chinese hedge fund also co-founded by…

DeepSeek V4.1 Flash Uncensored is a third-party "abliterated" build of DeepSeek's V4.1 Flash, published by the dealignai team rather than DeepSeek itself. The base model is a multimodal Mixture-of-Experts system with 552B backbone parameters that natively ingests images and text and generates text autoregressively, with continuously controllable reasoning effort. The fork applies weight-level edits to a small number of tensors to remove refusal behavior, while the routed experts, vision encoder, router gates, embeddings and speculative-decoding draft head are left untouched.

Relative to its same-family predecessors, the V4.1 generation differs architecturally from DeepSeek V4 Flash 0731 and DeepSeek V4 Flash 0423 by adding a from-scratch DeepSeek-ViT encoder for native multimodality and a long-context design supporting up to one million tokens, per the official V4.1 Flash card. The uncensored variant tracks the stock DeepSeek V4.1 Flash release of September 2026; a parallel abliteration also exists for the refreshed DeepSeek v4.1 Flash (1017) checkpoint.

Capability claims for the fork derive from the base model. On the published V4.1 Flash card figures mirrored by NVIDIA, the instruct model reports 90.9 GPQA Diamond pass@1, 90.6 on Terminal-Bench 2.1, and 74.2 resolved on DeepSWE v1.1. Abliteration is lossy by nature, so treat those as upper bounds for this variant.

Practically, it serves as a drop-in replacement for anyone already running the base checkpoint — same FP8 format, same loaders — but with safety declines removed, which shifts responsibility for content policy entirely onto the deployer.