Llama abliterated, hosted

The list is short. As of November 2026 the catalog carries abliterated builds on Llama 3 70B and Llama 3.3 70B, and an abliterated DeepSeek R1 distill on Llama 70B weights. The distill is listed here because it is a Llama model underneath, trained to reason in the R1 style; it also appears on the DeepSeek page.

All are tagged abliterated. There is no cybersecurity-tuned Llama build and no newer Llama generation in the list, and we would sooner say so than pad it. They are community builds on Meta’s weights, not Meta releases. The table on this page is read from the live catalog, including each build’s context window, and a build’s model page shows its current availability.

Llama 3 is the model on which abliteration became widely known. In 2024 researchers showed that refusals in chat models are governed largely by a single direction in the model’s activations, and community authors applied that to Llama 3: measure the direction on prompts the model declines and prompts it answers, then project it out of the weights. No retraining is involved, so the model keeps its knowledge and its style, and stops declining by reflex. What is abliteration explains the method step by step.

That history is why the Llama builds are a useful reference. Many public write-ups and comparisons were made against them, so if you are reproducing someone else’s results, or want a well-understood baseline to compare a newer build against, these are the ones to call. Abliteration still means reduced refusals, not zero.

A 70B Llama is a solid general model. It writes clear prose, follows instructions closely and explains concepts well, which covers a lot of security work: explaining a CVE advisory, wording a finding and its remediation, mapping results to CWE or MITRE ATT&CK, and drafting an executive summary.

It is also an older generation. The context windows on these rows are short by current standards, so a full scan export or a large source file will not fit in one request, and newer families are stronger at code review and multi-step reasoning. The R1 distill adds a visible reasoning pass, which helps on CTF problems that need several steps, at the cost of more output tokens.

As of November 2026 none of the Llama rows publishes tool calling; the R1 distill publishes a reasoning parameter and nothing else. Use them for chat-shaped work. For an agent loop that calls tools, choose a build whose page on the models list shows tools, or start from the router id auto/security.

Every row lists the OpenAI-format chat completions endpoint and the Anthropic-format Messages endpoint, so the curl request at the top of this page works once you replace auto/security with a Llama id from the table. If you already point a tool at a local OpenAI-compatible server, the change is a base URL, a key and a model id; the guides for OpenCode, Claude Code, Cursor, Cline and aider show where each goes.

The abliterated Llama weights are on Hugging Face and plenty of people run them. A 70B model at 4-bit quantization needs roughly 40 GB of GPU memory before context, which in practice means two 24 GB cards or one workstation-class card. Squeezing it onto a single consumer card means a smaller quantization and a real drop in quality, or spilling into system memory and waiting.

If you own that hardware and the work must stay offline, run it locally. If you do not, the arithmetic is simple: the cards cost money whether or not you are on an engagement, and an API costs nothing while idle. You also skip the driver, runtime and quantization upkeep. The trade is that prompts leave your machine, so check what your rules of engagement allow.

For most new work a newer family will serve you better: Qwen and GLM for cybersecurity-tuned builds, long context and tool calls, and Gemma if you want a compact model. Keep Llama for baselines, for prompts you have already tuned against it, and for plain writing tasks. The acceptable use policy applies throughout: lawful work on systems you own or have written permission to test.