This is a simplified guide to an AI model called Deepseek-V4-Pro-Qwen3.5-9b-Uncensored-Heretic maintained by Xtools-At. If you like these kinds of analysis, you should join AImodels.fyi or follow us on Twitter.

Overview

xtools-at presents this as a decensored version of Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B, produced with Heretic v2.0.0.dev0 using Arbitrary-Rank Ablation. The underlying 9B model is a Qwen3.5-9B reasoning fine-tune distilled from DeepSeek-V4-Pro, trained primarily for mathematics and STEM; this exact Heretic variant reports 3 refusals in 100 tests and KL divergence of 0.0422 against its original.

Best use cases

Mathematical problem solving, step-by-step derivation, physics and chemistry questions, broader STEM reasoning, and analytical instruction following. The underlying release also reports preliminary small gains in programming and tool-calling despite no coding data in training; this is not evidence of coding specialization.

Limitations

Coding and agentic coverage is limited, and the reported programming/tool-use transfer is preliminary rather than a substitute for dedicated benchmarks. The refusal and KL figures are reported for this Heretic variant, but the source does not describe the test prompts or methodology in detail. Benchmark results for the underlying model are not explicitly re-evaluated on this exact decensored variant.

How it compares

The source compares this variant with its original model: 3/100 versus 100/100 refusals, and KL divergence 0.0422 versus 0 by definition. For related model listings, see DeepSeek-V4-Pro-Qwen3.5-9B-MTP-GGUF, DeepSeek-V4-Pro-Qwen3.5-4B-MTP-GGUF, and Qwen3.8-27B-heretic-ara. These are separate model listings; the supplied evidence does not establish a quality, speed, or cost ranking among them.

Technical specifications

The exact variant is described as a Heretic v2.0.0.dev0 ARA decensoring of the 9B base fine-tune. ARA settings: start layer 9, end layer 24, preserve-good-behavior weight 0.8264, steer-bad-behavior weight 0.0003, overcorrect-relative weight 0.9645, and neighbor count 12. The underlying model was trained with approximately 250,000 math and STEM samples using SFT followed by light GSPO reinforcement-learning refinement; no coding data was included. Metadata lists Apache-2.0, Transformers, safetensors, and text-generation support. The source reports underlying-model evaluations using vLLM NVFP4 and llama.cpp MTP-Q8_0 GGUF, but does not establish that those benchmark scores apply to this exact Heretic variant.

Model inputs and outputs

Inputs

Text prompts for text generation and conversational use are supported by the supplied metadata. No exact prompt template, token limit, or generation-control settings are provided.

Outputs

Text generation is the listed task. The underlying release reports strong output-format compliance in its own evaluation, but those results are not established for this exact Heretic variant.

Frequently asked questions

Q: What distinguishes this exact model from the underlying 9B release?

A: It is described as a decensored version made with Heretic v2.0.0.dev0 and ARA. The source reports 3 refusals in 100 tests and KL divergence of 0.0422 against the original.

Q: What is the model intended to do?

A: Its underlying training emphasizes mathematical reasoning, derivation, physics, chemistry, and other STEM questions. It is also presented for structured analytical instruction following.

Q: Was it trained specifically for coding?

A: No coding data was included in the described SFT or RL stages. The source reports preliminary small programming and tool-calling gains as cross-domain transfer.

Q: Do the listed GSM8K and MMLU-Pro scores apply to this Heretic variant?

A: The source gives those scores for the underlying DeepSeek-V4-Pro-Qwen3.5-9B release and does not say that the exact decensored variant was evaluated on those benchmarks.

Q: What formats or runtimes are documented?

A: Metadata lists Transformers and safetensors. The underlying release was evaluated with vLLM NVFP4 and llama.cpp MTP-Q8_0 GGUF, but the source does not confirm those evaluation builds for this exact variant.

Q: What license is listed?

A: The supplied metadata lists Apache-2.0.

Click here to read the full guide to Deepseek-V4-Pro-Qwen3.5-9b-Uncensored-Heretic