language:

• en
• es
• fr
• ru
• it
• th
• zh
  license: other
  license_name: deepseek-v41-flash-1017
  license_link: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017/blob/main/LICENSE
  base_model: jangq/DeepSeek-V4.1-Flash-1017-JANG_2K
  library_name: mlx
  pipeline_tag: image-text-to-text
  tags:
• mlx
• jang
• jang_2k
• deepseek
• deepseek-v41
• deepseek_v41
• moe
• quantized
• apple-silicon
• 256gb
• reasoning
• thinking
• agent
• tool-use
• vision
• speculative-decoding
• dspark
• uncensored
• abliterated
• harmbench
• mmlu
• 1m-context
  thumbnail: dealign_mascot.png

DeepSeek-V4.1-Flash-1017 — UNCENSORED

JANG_2K · 256 GB Macs · 1M context

Uncensored across 7 languages (en / es / fr / ru / it / th / zh) · Chat + thinking (low / high / max) · Image input, tool use, DSpark speculative decoding · Everything the base bundle does, without the refusals

What this is

This is jangq/DeepSeek-V4.1-Flash-1017-JANG_2K with refusal behavior removed. Loads and serves exactly like the base bundle. The model answers questions it would otherwise decline, in every tested language and reasoning tier.

Results on this exact bundle

MMLU — 5700 questions, 100 per subject across all 57 MMLU subjects, chat mode

HarmBench-320 — standard test subset, chat mode (greedy T = 0, max_tokens = 512)

Real-harm ASR (ex-copyright, N = 240): 240 / 240 = 100.00 %.

Multi-turn pressure (chat, 8 harsh HarmBench-style prompts in one growing session)

Reasoning tiers

One uncensoring; all four reasoning tiers.

Note on greedy + max-effort reasoning: at effort = max with greedy sampling and growing context, DeepSeek-V4.1 can get stuck in self-correcting knowledge-recall ruts on hard structural prompts. This happens on the base bundle too. Serve with the recommended temperature = 1.0, top_p = 1.0 (see generation_config.json); temperature noise breaks these ruts.

Serving

Load it exactly like the base bundle: Osaurus 0.25.27+, or vmlx serve with any recent jang_tools build that supports this family. Apple Silicon ≥ 256 GB.

Thinking mode via the bundle's own chat template:

"template_vars": { "enable_thinking": true, "reasoning_effort": "high" }

reasoning_effort accepts "low", "high", "max", or an int 1..100. Chat mode: { "enable_thinking": false }.

Serving sampling defaults (generation_config.json): temperature 1.0, top_p 1.0 (per the DeepSeek-V4.1 recommendation — kept unchanged from the base bundle).

License & credits

• Base model & license: deepseek-ai/DeepSeek-V4.1-Flash-1017, under the DeepSeek V4.1 Flash license.
• JANG_2K bundle: jangq/DeepSeek-V4.1-Flash-1017-JANG_2K. Thanks to the JANG team for the Apple-Silicon pack this release is built on.
• Uncensoring: dealignai.

[!NOTE]
Please credit if you build on this. If you redistribute, fine-tune, serve, benchmark, or write about this model, a line like "Based on dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-JANG_2K" in your model card means a lot. Thank you.

Safety

This model has refusal behavior removed. It will provide information it would otherwise decline. You remain fully responsible for how you use it. Do not use it to harm real people.