A disclaimer is not an answer.
A great deal of legitimate research runs straight into content policy. Journalists reconstruct how a crime worked. Security researchers need to understand an attack before they can defend against it. Clinicians and harm-reduction workers need specifics, not general wellness advice. Historians, policy analysts and novelists all need a model that will describe atrocity in the terms the subject requires. Mainstream assistants are calibrated for hundreds of millions of users at once, which means they refuse a large amount of work that is entirely lawful and entirely serious.
OpenRogue's lineup is calibrated differently. Nous Research trains the Hermes models on an explicitly neutral alignment philosophy — the user is treated as the authority on what they need — and Dolphin was fine-tuned to answer rather than to moralize. Combined with 131K-context models that can read a stack of documents in one sitting and a Thinking mode for deliberate analysis, that makes for a genuinely usable research assistant. It also makes verification your job, and that part is not optional.
Research on difficult topics fails at the first step if the model will not discuss the topic. Half-answers with the specifics removed are worse than nothing, because they look like answers.
Real analysis means holding several documents at once and reasoning across them. The 131K-context models take a transcript, a filing and a report together and compare them rather than summarizing each in isolation.
Summarizing is the easy half. The value is in a model that will weigh two contradictory sources, say which is more credible and explain why — which requires a model willing to commit to a judgement.
Some questions need the model to work rather than answer. Thinking mode plus a hybrid reasoner like Hermes 4 produces visibly different quality on multi-step analysis than an instant reply does.
The most important thing a research assistant can say is 'I am not sure'. You get closer to that by asking for confidence levels and for the strongest counter-evidence, rather than accepting the first fluent paragraph.
The subject itself is declined, regardless of purpose. Security, pharmacology, extremism, weapons policy, historical atrocity — entire fields where the assistant will not engage, and stating your professional context rarely changes the outcome.
More common and more insidious than a refusal: you get a fluent, well-structured response that has been stripped of every specific that would make it useful. It reads as an answer, so you may not notice what is missing until you try to use it.
On anything contested, safety-tuned models present a balanced summary and decline to weigh the evidence. For a researcher that is the opposite of help — you came for a judgement, and you get a list.
Paragraphs of caveat wrapped around a two-sentence answer, plus a suggestion to consult a professional you have already consulted. Individually trivial, collectively a real tax on any session longer than a few questions.
| Model | Role | Why |
|---|---|---|
| Hermes 4 | Analysis with visible reasoning | A hybrid reasoner built for exactly this: turn Thinking mode on and it works multi-step problems deliberately instead of pattern-matching an answer. Nous's neutral alignment means it argues positions on the merits rather than retreating to balance. |
| Hermes 3 405B | Depth and nuance | Frontier scale with 131K context, for questions where the answer depends on holding a lot of world knowledge and several documents at once. It is the model for ethically complicated questions that smaller models flatten into platitudes. |
| WizardLM-2 | Technical explanation and structure | Its specialty is long multi-constraint instruction following, which is what a proper literature summary or a comparative technical breakdown actually is. It produces long structured documents without losing the format or the thread. |
| Dolphin Venice | Fast direct answers | Trained to answer rather than to hedge, and quick enough for the rapid question-and-follow-up rhythm that most research actually has. It is the free-tier model, which makes it a reasonable place to test whether a line of enquiry is worth pursuing. |
This is the use case where the honest warnings matter most. Removing a refusal layer does not make a model more accurate — it makes it more willing, and willing plus wrong is a genuinely dangerous combination. These models hallucinate citations, misattribute quotations, and state dates and figures with complete confidence when they are simply wrong; a fabricated reference looks exactly like a real one. Web search is available as a toggle on paid plans and it bills per result, but it retrieves rather than verifies, and the model can still misread what it retrieved. Training data has a cutoff, so recent events are unreliable by default. And an uncensored model is a research tool, not a licensed professional: for medical, legal or financial decisions it can help you understand a subject and prepare better questions, and it cannot replace someone accountable for the advice.
/// Standard
Specifics over generalities. No disclaimers. Flag every claim you are unsure of.
/// Compare
These two sources contradict each other. Which is more credible, and why?
/// Counter-case
Now argue the strongest case against what you just told me.
No — and conflating the two is the most common mistake people make here. Uncensored means the model will engage with a subject rather than decline it; it says nothing about whether the answer is right. Every load-bearing fact, citation and figure still needs checking against a primary source.
The pattern clusters around security and exploitation, pharmacology and harm reduction, extremism and radicalisation, weapons and conflict, explicit historical atrocity, and specifics in medicine and law. All are ordinary subjects of journalism, scholarship and professional practice, and all routinely trigger refusals or answers with the substance removed.
Yes — file understanding and vision input are available on paid plans, and the 131K-context models can hold a substantial stack of material at once. Working from documents you supply is markedly more reliable than asking the model to recall a source from training, so upload rather than ask wherever you can.
There is a web search toggle on paid plans. It is off by default, it bills per result on top of tokens, and it retrieves sources rather than verifying them — the model can still misread or over-claim from what it finds. Treat retrieved results as leads to check, not as citations.
OpenRogue does not train on your prompts and does not sell your conversations to partners. Your history is synced to your account, and you can export it or delete it. If you are handling material with a legal confidentiality obligation, check that a hosted service is permitted under whatever agreement covers it.