Model cards: Arlo, Lucian, Orion and Curzon

Last updated 5 November 2026. Artemii is made by Mayfair.

These cards describe Artemii's four language models in plain terms. None of these models has been independently benchmarked.

Common to all four

What they are

Arlo, Lucian, Orion and Curzon are Artemii's language models. Each is built on an open-weight AI model that Mayfair adapts and configures, combined with Mayfair's own instructions, tools, safety settings and software. They are different configurations tuned for different kinds of work, balancing depth of thinking against speed. Mayfair did not train the underlying model from scratch, and "adapts" here means configuration, instructions, tools and settings. Credits, licence notices and the creators' own usage terms are on the Licences and model notices page.

Where they run

On GPU servers that Mayfair rents and controls (see Sub-processors). Your messages are not sent to a third-party AI service to be answered.

Intended use

General assistance for adults and for working organisations: conversation, writing, summarising, analysis, planning, coding and design, creating documents and web pages, research with the web tools, and working with your files and connected services. Not intended as the sole decision-maker for anything with legal or similarly significant effect on a person, and not a substitute for a lawyer, doctor, financial adviser or other professional. General use of Artemii is for people aged 16 or over; Direct mode is for people aged 18 or over.

Out of scope

See the Acceptable Use Policy. In particular: sexual content involving children, weapons capable of mass harm, malware, targeting people, deception, and decisions about people without human review.

Inputs and outputs

Input: text, the text of text files, and tool results. The app stores uploaded pictures, but the models do not read picture content and say so. Output: text (Markdown), documents, web pages and designs (artifacts), code, spreadsheets, and requests to use tools. Pictures are made by a separate picture model, not by these four. The developer API accepts text only. The interface is in English, and no language-by-language testing has been recorded.

Limits enforced in code

Known limitations (all models)

• They generate text statistically and can be wrong, out of date, incomplete, inconsistent or biased. They can state falsehoods confidently, including invented facts, quotes, sources or links, although their instructions tell them not to.
• They have no live knowledge unless web tools are on and used, and they are told not to pretend to have searched.
• They can be steered by malicious text in web pages, files or tool results (prompt injection). The product wraps such text as untrusted data and asks you before actions that change something, but this does not make a model immune.
• They do not remember between chats except through the memory feature you control.
• Arithmetic, long multi-step logic, rare languages, niche domains and very long documents are likely weaker areas. This has not been measured.
• Output can resemble other people's text or code and is not guaranteed free of third-party rights.
• Safety behaviour is an AI model's behaviour, not a technical guarantee. The only hard filter before the model is a keyword screen on picture prompts for sexual content involving minors and sexual images of real people, and it is a heuristic.

Built-in behaviours

The models present themselves as Artemii (under the model's name), made by Mayfair; never claim to be human and say they are an AI if sincerely asked; do not invent specifics about themselves; treat text from the web, files and tools as data rather than instructions; and follow hard limits (no help with mass-destruction weapons or child sexualisation, no working weapon, explosive or drug-synthesis instructions, no malware or attacks on others, no targeted harassment, stalking, doxxing or fraud). In Direct mode (adults only, opt-in) the tone is more candid with fewer disclaimers; the limits stay.

Evaluation status

• Not independently benchmarked. No third-party or published benchmark results exist for these models.
• Internally, a set of identity and behaviour checks confirms with simple keyword rules that a model gives its Artemii name, says it is an AI, does not leak its instructions and keeps a plain style. Keyword checks catch obvious failures only.
• No red-team, bias or fairness evaluation, hallucination-rate measurement or language-by-language testing is recorded.
• Thumbs up and down are counted per model per day, anonymously and without content. They are not a benchmark.

Safety and oversight

Hard limits in the instructions; actions that change something ask first; permissions; Direct mode behind an adult acknowledgement; rate limits; administrator controls. You can stop any answer with the Stop button.

Data used by the model

A prompt contains Artemii's instructions, your custom instructions, memory notes (most relevant first), project instructions and files, attached files, the chat history and tool results. Content you have not added to a chat is not in the prompt. The training data of the underlying model is described by its creators; Mayfair does not hold or control it.

Arlo (version 3.2)

Quick answers and conversation.

• What it is for: Chat, short questions, rewrites, simple look-ups and summaries. Our fastest and friendliest model, and the default.
• How it behaves: Answers straight away with light thinking. Leads with the answer and keeps replies short and warm. For work that clearly needs long, careful effort it does its best briefly and mentions that Lucian is built for harder problems.
• Use with care for: hard maths, multi-step logic, and anything where a thin answer would cause harm.
• Speed and depth: fastest, shallowest.

Lucian (version 3.5)

Hard problem solving.

• What it is for: Difficult maths, logic, planning and multi-step reasoning.
• How it behaves: High reasoning effort. Takes longer, works step by step, checks its reasoning and arithmetic, considers alternatives, says plainly when something is uncertain, and gives a clear final answer with the key steps.
• Use with care for: its reasoning and answer can still be wrong, so verify results that matter. It uses more of your allowance than Arlo.
• Speed and depth: slowest, deepest.

Orion (version 3.0)

Research, OSINT and data analysis.

• What it is for: Analysing sources, documents and data; structuring findings; briefings, timelines and reports; lawful open-source research.
• How it behaves: High reasoning effort. Works like a careful analyst: separates facts from inferences, states confidence, says what is unknown, flags gaps, bias and conflicting sources, and puts the summary first. By default it does not browse unless web tools are on and requested, it says so plainly when live web access is unavailable, and it never invents sources, quotes or links. It can offer deep research when a question looks like research and you have runs left.
• Use with care for: research about named individuals. Research must be lawful and the Acceptable Use Policy applies. Output is a lead, not evidence, so check sources. Not for decisions about people without human review.
• Speed and depth: slow, deep.

Curzon (version 3.1)

Design and coding.

• What it is for: Writing, fixing, explaining and reviewing code; building web pages, apps, scripts and interface designs as artifacts.
• How it behaves: Medium reasoning effort. Writes clean, working, well-structured code and modern, accessible designs, prefers complete working artifacts to fragments, and explains its choices briefly.
• Use with care for: generated code can have bugs, security flaws and licence issues, so review and test before use. Generated pages run in a sandbox with no network access, and design accessibility is not checked.
• Speed and depth: medium.

About version numbers

The version numbers shown in the product (3.2, 3.5, 3.0 and 3.1) are product labels.

Questions: hello@mayfairmail.com. See also Models and AI transparency.