What happened
Hugging Face hosts a public, ungated repository named DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic. Its model card calls it a “decensored” version of Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B, created with Heretic v2.0.0.dev0 and its Arbitrary-Rank Ablation, or ARA, method. The repository is tagged for text generation and image-text-to-text use, has an Apache 2.0 licence tag, and lists English, Chinese, Korean, Japanese, Spanish, and Russian among its languages. At the time shown on the page, it had 7,088 downloads and 774 likes.
The card identifies the modified checkpoint’s starting model as DeepSeek-V4-Pro-Qwen3.5-9B, itself described as a reasoning-focused fine-tune of Qwen3.5-9B distilled from DeepSeek-V4-Pro responses in “Max Effect” mode. It says that underlying model was trained on approximately 250,000 mathematics and STEM samples. The training path is described as supervised fine-tuning followed by a light reinforcement-learning refinement using GSPO, short for Grouped Stepwise Preference Optimization. The publisher says no coding data was included, while later tests showed small improvements in programming and tool-calling tasks; that is a publisher claim about the underlying training line.
For the Heretic alteration, the card lists ARA settings: start layer index 9, end layer index 24, preserve-good-behavior weight 0.8264, steer-bad-behavior weight 0.0003, overcorrect-relative weight 0.9645, and neighbor count 12. Its comparison table reports 3 refusals from 100 prompts for the altered model and 100 refusals from 100 prompts for the original checkpoint. It also reports a KL divergence of 0.0422 for the altered checkpoint, with zero for the original by definition. These figures describe the publisher’s own comparison rather than a replicated external result.
The card separately reports benchmark results for DeepSeek-V4-Pro-Qwen3.5-9B: a 94.50% four-run GSM8K average under vLLM, plus a 90.53% average over 1,500 MMLU-Pro mathematics, physics, and chemistry questions. Its MMLU-Pro table compares an MTP-Q8_0 GGUF build at 8K context with Qwen3.5-9B and two Claude Mythos-distilled references. The card also reports 36.1% fewer tokens per correct answer than Qwen3.5-9B in that test. Those underlying-model benchmarks should not be treated as an evaluation of the refusal-modified checkpoint.
The repository lists 9,409,813,744 total parameters and 19,306,374,364 bytes of safetensors files. Hugging Face’s usage panel includes examples for Transformers, vLLM, SGLang, and Docker Model Runner. The vLLM and SGLang examples expose OpenAI-compatible chat-completions endpoints, while the Transformers example loads the model with AutoProcessor and AutoModelForMultimodalLM. The page also directs users to quantizations for llama.cpp, Ollama, LM Studio, or compatible applications.
What the archive adds
New Runtime covered a separate community Qwen3.5-4B-heretic release on November 1, 2026, described as sharply reducing refusals with little reported benchmark drift. This is a new 9B checkpoint in the same community-modification thread, rather than the same model or demonstrated continuation of that release.
Trend signal
This supports no specific pattern New Runtime tracks. The notable change is a checkpoint-level alteration to refusal behavior, not a new agent runtime, protocol, capability interface, or verification mechanism. It may become more consequential if such modifications gain reproducible evaluations and operational deployment practices, but this individual release does not establish that broader shift.
Why it matters
For teams running open-weight models, refusal behavior is part of the deployable surface. It determines which requests a system will attempt, but it does not determine whether those answers are correct, safe, compliant with policy, or useful in a workflow. A reported shift from 100 refusals to three therefore warrants attention from teams whose local model rejects required benign tasks.
It should also change the evaluation question. The relevant comparison is not simply whether an altered checkpoint refuses less often. Teams need to compare task completion, factual accuracy, format compliance, tool use, latency, token use, and behavior on requests that should remain blocked. A model that is more permissive may reduce friction in one workflow while increasing the need for application-level controls, review, and monitoring.
Limits and uncertainty
The refusal result is a publisher-run test of 100 prompts. The card reports the counts, but the result has not been independently replicated or independently reviewed here. A refusal count alone cannot measure answer quality, harmful compliance, robustness, language coverage, or production reliability. Likewise, the benchmark numbers, token-efficiency calculations, training description, and claimed programming or tool-use transfer are publisher-reported.
The model card distinguishes benchmark tables for the underlying distilled checkpoint from the decensored variant, so capability claims should not be transferred automatically between them. Hardware requirements in practice will vary by quantization, context length, runtime, batch size, and serving configuration. The availability of integration examples is not evidence that each runtime produces equivalent outputs or performance.
Skill to build now
Create a small evaluation harness before trialling an altered local checkpoint. Include representative requests your application needs completed, requests that should be rejected, structured-output checks, and a human review sample. Run the base and modified models with the same prompt template, decoding settings, hardware, and runtime. Record refusal rate separately from correctness and unsafe compliance; otherwise a lower refusal number can conceal a worse operational result.
Reader takeaway
This is a public community modification of a 9B Qwen-derived checkpoint, not a new foundation model release. Its reported refusal reduction is striking, but it is only a starting hypothesis for local evaluation. Treat permissiveness as one measurable model property and test it alongside quality, safety boundaries, and the controls your application supplies.