This DeepSeek v4.1 is powerful and uncensored, let's run it

Summary

David Ondrej explains what he calls “obliterated” or uncensored models: open-weight language models modified to reduce or remove refusal behavior. He argues that they can help with legitimate tasks that filtered models may decline, including security testing, sensitive-topic research, and other ordinary requests. However, he stresses that they can also produce harmful guidance or take destructive actions if connected to tools, and advises using them legally and carefully.

How refusal removal works

The video outlines a model pipeline of pretraining, instruction fine-tuning, and safety training such as RLHF. The presenter’s explanation is that refusals are partly a behavior layered onto a model, rather than a lack of underlying knowledge. He describes one modification approach as identifying and changing model-weight directions associated with refusals. He also cautions that this works only with models whose weights are available, not closed services such as ChatGPT or Claude.

The video compares three approaches:

1. Jailbreaking: Alters prompts to get a model to ignore restrictions. It is quick to try, but increasingly unreliable and easy for providers to patch.
2. Fine-tuning: Trains the model on additional examples. It can change response behavior, but requires data and may affect capability.
3. Weight modification (“abliteration”): Changes selected weights to reduce refusals while attempting to preserve the model’s other capabilities.

The presenter describes Heretic, an open-source GitHub tool, as an automated way to perform this kind of modification without requiring users to understand transformer internals. He claims the changes can substantially reduce refusals with little measurable loss in benchmark performance. These are claims made in the video and have not been independently verified here.

Models and examples mentioned

• The presenter points viewers to Hugging Face and ModelScope for open-weight models, including modified versions.
• He mentions HuiHui AI as a Hugging Face publisher with multiple modified models, and Junsong in connection with “Super DeepSeek” and “Super Gemma.”
• Model names in the subtitles include Qwen 3.8 27B, Super DeepSeek V4 Flash, and other “obliterated” releases. Some names or version numbers may be inaccurate because the subtitles are auto-generated.
• In a demo, the presenter runs a large DeepSeek-based model through an API-backed setup and asks it general and cybersecurity-related questions. He warns that models connected to tools can carry out destructive actions without safeguards users may expect.

Running models: guide and hardware advice

The presenter gives three deployment options:

• Third-party API: Easiest way to access large models without owning hardware, but it costs per token and sends requests to an external provider.
• Local computer: Can offer offline, private use without per-token fees, but is limited by available memory and processing speed.
• Rented cloud GPUs: Can run larger models and are billed by time, but require more technical setup and are not as private as running locally.

For local use, the video demonstrates Ollama and describes other options such as LM Studio, llama.cpp, and vLLM. The basic walkthrough is to find a model, check that its format is supported—particularly GGUF for some Hugging Face imports—and run it through Ollama. The presenter also notes that an Ollama terminal session can be inconvenient without a separate chat or coding interface.

The hardware guidance is presented as rough estimates that depend on model size, quantization, and memory bandwidth. The speaker gives a rule of thumb of roughly 0.6 GB of memory per billion parameters and says Apple Silicon’s unified memory can be used by both the CPU and GPU. He offers approximate model-size ranges for systems with 8, 16, 32, and 64 GB of memory, while noting that the figures are not guarantees. Larger models may require multiple GPUs or high-memory systems. He also emphasizes that memory bandwidth affects generation speed.

The presenter recommends downloading and archiving open model weights, arguing that models or hosting platforms could become unavailable. This is his advocacy and prediction, rather than a technical requirement for running a model.

Security and broader analysis

A major proposed use case is defensive security testing: running less-restricted models against one’s own software to identify weaknesses that more heavily filtered models might not explore. The presenter argues that organizations should use both capable mainstream models and less-restricted models for this purpose. He also cites research and recent incidents to support his case for reduced refusals, but the subtitles do not provide enough detail to validate those references.

He argues that less-restricted open models may be useful even if they perform below leading closed models on some benchmarks, because they may comply with more requests. The video also promotes saving open model weights and describes them as important for access to AI.

Product and service mentions

• PostHog (sponsor): Product analytics and session replay for understanding user behavior, feature use, and retention.
• Cloudroom: The presenter’s open-source coding environment, which he says supports multiple model providers and operating systems. He demonstrates using it with a model served by his local hardware.
• Venice: Mentioned as an API option that the presenter says does not require KYC. He advises checking the privacy and identity-verification terms of any service.

Main speaker and sources

• Main speaker: David Ondrej (introduced in the subtitles as “David Andre”).
• Tools and sources discussed: Heretic (GitHub), Hugging Face, ModelScope, Ollama, LM Studio, llama.cpp, vLLM, Venice, Cloudroom, and PostHog.
• Model creators or publishers mentioned: HuiHui AI and Junsong.