DeepSeek-V4-Pro-Qwen3.5-9B Uncensored Heretic Explained (2026)
DeepSeek-V4-Pro-Qwen3.5-9B Uncensored Heretic Explained (2026)
Short answer: Hugging Face repo xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic is a community decensored build of Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B. The base is a 9B Qwen3.5 student distilled from DeepSeek-V4-Pro (Max Effect) on ~250K math/STEM samples. The Heretic build then applies Arbitrary-Rank Ablation (ARA) so reported refusals drop from 100/100 → 3/100 at KL 0.0422. It is local open weights, not a hosted DeepSeek product. If you want creative chat without a GPU desk, use a cloud workspace such as HackAIGC Chat.
This page explains the model card and the local-vs-cloud choice. It is not a duplicate of DeepSeek V4 Pro NSFW / uncensored product pages or the V4.1 Flash abliterated GGUF explainer (different base + different packaging).
TL;DR
| Fact | Detail (from the public model card, Nov 2026) |
|---|---|
| Repo | xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic |
| Base student | Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B — 9B, Qwen3.5 backbone |
| Teacher | DeepSeek-V4-Pro in Max Effect mode (~250K math/STEM distillation samples) |
| Decensor tool | Heretic v2.0.0.dev0, ARA on layers 9–24 |
| Refusal metric | 3/100 vs 100/100 on the original student |
| KL divergence | 0.0422 (small distribution shift; not a full retrain) |
| Size class | ~9B BF16 (~18GB class); community GGUF and MLX quants exist |
| Best for | Local tinkerers who want that exact Heretic label |
| No-GPU path | Hosted uncensored chat — https://chat.hackaigc.com/ |
One-sentence definitions
| Phrase | Plain English |
|---|---|
| DeepSeek-V4-Pro-Qwen3.5-9B | A 9B reasoning student distilled from DeepSeek-V4-Pro into a Qwen3.5-9B backbone (math/STEM-heavy) |
| Heretic / UNCENSORED-heretic | A post-training refusal edit made with the Heretic tool (ARA), shipped as a new HF checkpoint |
| ARA (Arbitrary-Rank Ablation) | Weight surgery that subtracts a measured refusal component inside a layer range |
| KL 0.0422 | How far the edited model’s token distribution drifted from the student — low here means “brake removed,” not “model rebuilt” |
| GGUF / MLX quants | Community re-packages so you can run the same lineage on llama.cpp / Apple Silicon stacks |
For the broader vocabulary (abliterated vs uncensored vs no-filter), see What Is an Abliterated Model?.
Three stages behind the trending name
Think of the HF title as a pipeline, not a single official DeepSeek SKU:
- Teacher — DeepSeek-V4-Pro (Max Effect) produces math/STEM solutions.
- Student — Those traces distill into Qwen3.5-9B →
Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B(SFT + light GSPO RL; no coding SFT in the published mix). - Heretic edit —
xtools-at/…-UNCENSORED-hereticremoves refusal behavior with ARA.
Stages 1–2 explain the reasoning / STEM pitch. Stage 3 explains why the repo is labeled uncensored and why it trended next to other refusal-edit releases.
What the student was built for (before Heretic)
Per the base card mirrored on the Heretic page:
- ~250K mathematics and STEM samples from DeepSeek-V4-Pro
- Reported GSM8K ~94.50% (NVFP4 / vLLM average across four runs)
- Reported MMLU-Pro Math/Physics/Chemistry average ~90.53% on a 500-per-subset sample (not full subsets)
- Strong format compliance claims on single-letter answer instructions
- Explicit limitations: experimental 9B, hallucinations remain, coding gains are preliminary transfer, not a coding specialist
Heretic does not republish a full before/after STEM re-benchmark on the edited weights — treat capability as “student-class until you re-test your own workload.”
How Heretic ARA changed refusals
Public ARA parameters on the Heretic card:
| Parameter | Value |
|---|---|
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 |
neighbor_count | 12 |
| Metric | Heretic build | Original student |
|---|---|---|
| Refusals | 3/100 | 100/100 |
| KL divergence | 0.0422 | 0 (by definition) |
Plain English: someone measured a refusal-related direction in the mid/late layers, subtracted it, and saved new weights. Refusals collapse. The distribution barely moves (KL ~0.04), which is why people call it weight surgery rather than a new fine-tune.
Related technique family (different tool names, same idea space): abliterated models.
Local weights vs hosted chat (no GPU)
| Dimension | Local Heretic 9B | Hosted creative chat (e.g. HackAIGC) |
|---|---|---|
| Setup | Download BF16 / GGUF / MLX; pick runner | Open a browser URL |
| Hardware | ~18GB BF16 class; ~6–10GB with 4/8-bit; GGUF Q4 often consumer-friendly | Provider GPUs |
| Label fidelity | Exact …-UNCENSORED-heretic checkpoint | Product says uncensored / no-filter — not a claim it is this HF file |
| Control | Samplers, system prompts, offline possible | Product UX + AUP |
| Modalities | Text (unless you bolt on other tools) | Chat + image + video in one stack |
| Best for | Research, red-team, offline, exact weights | Shipping creative drafts today without a GPU desk |
Rule of thumb: Care about method + exact Heretic weights → local HF. Care about shipping legal adult creative work today → HackAIGC Chat.
Community convenience packages (verify live on HF before you download):
- GGUF:
mradermacher/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic-GGUF - Apple Silicon MLX: community 4-bit conversions exist under separate uploaders
What this release does not guarantee
- Perfect removal of every hedge, deflection, or weird refusal remnant
- That STEM scores on the edited weights match the student card’s sampled benchmarks
- DeepSeek Corp blessing — this is a third-party Heretic of a community distill
- Legal cover for illegal prompts (law + host AUPs still apply)
- Privacy — local helps only if inference stays on your machine
- That a SaaS “uncensored DeepSeek” tab is this same file
Practical checklist if you open the HF page
- Read the model card (base attribution, ARA table, limitations).
- Confirm license notes on the student and the Heretic upload.
- Match file size to RAM/VRAM (BF16 vs Q4/Q5/Q8 GGUF vs MLX).
- Prefer known runners (Transformers, vLLM, llama.cpp, mlx-lm) over random binaries.
- Smoke-test your tasks — math, roleplay, tool use — after the refusal edit.
- If local ops are not worth it, keep prompts portable and draft in hosted chat.
Soft product bridge
HackAIGC markets a hosted uncensored / no-filter creative experience. That is a product promise, not a claim that the backend is xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic. Use local Heretic when you need that exact label; use HackAIGC Chat when you want zero install and multimodal creative work.
For hosted DeepSeek-shaped NSFW product intent (different page), see DeepSeek V4 Pro NSFW uncensored.
FAQ
Is DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic official DeepSeek?
No. The teacher is DeepSeek-V4-Pro; the student and Heretic builds are community releases on Hugging Face unless DeepSeek’s own org publishes them.
Is Heretic the same as abliteration?
Same family of idea (remove refusal direction), different tooling/branding. Heretic here uses ARA with the published layer range and weights above.
Do I need a GPU to “use Heretic”?
For this checkpoint, yes (or a strong CPU + quantized GGUF, still local). For uncensored creative chat without hardware, use a hosted product instead.
Heretic vs jailbreak?
Heretic/ARA = persistent weight edit. Jailbreak = prompt tricks against a still-aligned model.
Where should beginners start?
If you only need to write and iterate prompts: start hosted (chat.hackaigc.com). Graduate to local Heretic/GGUF when you care about offline control and the exact HF label.
Bottom line
DeepSeek-V4-Pro-Qwen3.5-9B Uncensored Heretic = a 9B DeepSeek-V4-Pro → Qwen3.5 distill, then refusal-ablated with Heretic ARA (100→3 refusals, KL 0.0422). It is a local weights / infra story — distinct from “NSFW DeepSeek hosted” marketing pages and from the V4.1 Flash abliterated GGUF line. Pick local Heretic for control; pick hosted chat for convenience.
— Ethan Cole