DumGumDocs
Search documentation

Loading search…

Guides

Choose Your Model

DumGum offers two model tiers: Smart and Pro. The difference is output quality, cost, and features. A third option, Auto, uses both in the same conversation: Smart first, then Pro once the conversation needs long-term memory.

Smart vs Pro

SmartPro
QualitySolid, reliable outputHighest quality output
Best forHigh-volume standard conversationsPremium or complex conversation flows
FeaturesSoulEngine, Core Guardian, Vision Link, Smart Share, Fictional Disclosure, Platform Information, User SuggestionsEverything in Smart, plus Long-term Memory
Cost€0.025 per reply€0.035 per reply

A request that names no model, or names default, runs on chat-2-smart, version 2.0. Always name the latest version instead: see Model Versions.

Both tiers give the Persona the same short-term memory: the 30 most recent messages of the conversation. What differs is what survives beyond that window: Pro adds Long-term Memory. See Conversation Memory for what actually occupies those 30 slots.

Auto: Smart first, then Pro

Each version also comes as an -auto model: chat-2.2-auto, chat-2.1-auto and chat-2-auto. It is not a third tier. It picks Smart or Pro for you, conversation by conversation, and it is the option we recommend.

Conversation lengthReply generated byLong-term MemoryBilled at
Up to 30 messagesSmartNot keptSmart rate
From the 31st message onProInitial messages analysed, then kept incrementallyPro rate

The message being replied to counts, so the first Pro reply is the reply to the 31st message. The tier is chosen again on every request from the conversation's message count, which only grows: once past 30, the conversation stays on Pro for as long as you keep sending the -auto model.

Why it switches at 30 messages

30 messages is the Persona's short-term memory window (see Conversation Memory). As long as a conversation fits in it, she sees all of it, so Smart is enough. The 31st message is where the oldest messages start leaving that window, and it is on that reply that Pro takes over.

Smart keeps no Long-term Memory. When Auto first moves a new conversation to Pro, it analyses the text in the first 30 and latest 31 messages, counting a message present in both groups only once. This includes the opening 30 even if split replies or queued messages make the conversation jump from 29 to 33 messages. Facts from those early exchanges can therefore become Long-term Memory even after they leave the reply's 30-message window. This analysis runs alongside reply generation, so its new memories may only be used by a later reply. Subsequent updates run once about 20 new messages have arrived.

Only Auto runs this catch-up: a conversation moved to a fixed Pro model by hand keeps no memory of what was said before its first Pro reply. See When Smart becomes Pro for the full rules, including longer conversations, retries and existing memory.

The User notices nothing: same Persona, same tone, same history. Each reply is billed at the tier that produced it, so a conversation that ends before its 31st message costs the Smart rate only.

When to pick Smart or Pro instead

  • Smart, if you never want to pay the Pro rate and can do without Long-term Memory.
  • Pro, if you want Long-term Memory as early as possible. It covers what is said after the conversation's first Pro reply: the opening message, and any history that existed before your first request, are not memorised.

Use it

Set replyParameters.chatModel to the -auto model of your version, for example chat-2.2-auto. The choice is made on each request; there is no project setting for it. A request that names a -smart or -pro model stays on that tier, whatever the length of the conversation. There is no chat-latest-auto.

Current and Earlier Versions

The current generation is chat-2.2-smart, chat-2.2-pro and chat-2.2-auto. They cost the same as the chat-2 models and support the same features, with the sharing and first-reply behaviour described in Control Sharing and Sexual Intensity and the question control described in Model Versions. Prefer chat-latest-smart or chat-latest-pro if you would rather follow the newest released version automatically: today they resolve to 2.2.

chat-2.1-* and chat-2-* remain available, with the features and settings they shipped with. That is deliberate: a pinned version never gains or loses a feature underneath you. We still fix safety issues and adjust how replies are produced on every version, without changing what your settings do. The flip side is that a setting introduced after your version is ignored there — replyParameters.questionFrequencyRate is read by 2.2 and later only, so if you send it to an earlier model nothing happens. All three generations support Vision and Smart Share features.

Already running on 2.1 or 2.0? Move to the latest version. It costs the same per reply, and your current version stays available while you switch. What each version changes, and what to check before switching, is listed in Model Versions.

Ask DumGum AIAnswers from the documentation

What can we help you build?

Ask about setup, content creation, or the API.

AI can make mistakes—verify important information against the sources. Conversations are kept for 30 days to improve the documentation, and questions and recent messages are sent to a third-party AI provider. Don’t include secrets or sensitive information.