Chat & Models

ETOS LLM Studio's edge isn't "you can send messages" — every LLM client does that. Its value is in how sessions become long-term assets, how tunable the parameters are, and how comfortable long AI responses are to read. This page covers all three.

For the actual order in which a message's context gets assembled, see Prompt & Context Assembly.

Read This First

Session Management as a Long-term Asset

A "session" is a continuous chain of messages. But ETOS treats them as assets you can categorize, search, share across devices, and branch — not just a transcript.

The Session List (top-left menu)

Tapping the menu on the chat screen opens the full session list:

• Search field with placeholder "Search session titles or messages" — does full-text message search, not just titles. Tapping a result jumps to the matching message.
• New conversation button
• Folders — group sessions by topic. Long-press a session → Move to Folder.
• Session entries — each shows title (auto-named or manually renamed) plus a preview of the last message.

What Long-press on a Message Does

Almost every per-message action lives in the long-press menu:

Action	Use
Copy	Copy message text
Quote	Quote this message in your next input
Edit	Rewrite either your input or the AI's reply
Delete	Remove this single message
Token Info	See prompt / completion / thinking tokens
Thinking Duration	Reasoning models only — how long the AI thought
Export Entire Session	Save the whole conversation
Export Up to This Message	Save up to (and including) this one
Create Prompt Branch	Spin off a new session from here (see below)

Branching

Long-press → Create Prompt Branch opens a "Branch Options" dialog with two choices:

• Prompt only — clone the system prompt, no messages
• Messages only — clone everything up to here (including AI replies)

The branch is a completely independent session; the original is untouched. Use it for:

• "Try the same question with a different tone"
• "Diverge here without polluting the main thread"
• "Use this thread as a template for a fresh start"

Multiple Providers and Models

ETOS supports four API formats (not just four vendors — one format covers countless vendors):

API Format	Internal ID	Who uses it
OpenAI Compatible (Chat Completions)	openai-compatible	Official OpenAI legacy + nearly every third-party relay + China-based compatible services
OpenAI Responses	openai-responses	OpenAI's newer endpoint for GPT-4.1 / GPT-5
Anthropic	anthropic	Claude official
Gemini	gemini	Google official

Key concepts:

• A provider has one API format — pick Anthropic when adding Claude.
• A provider can host many models — the same OpenAI key works for GPT-4o, GPT-4o-mini, text-embedding-3-large, whisper-1, tts-1 simultaneously.
• Models are slotted by capability: chat / image / embedding / rerank / speech-to-text / text-to-speech. A model can have multiple capabilities.

Assigning Models by Capability

In Settings → Providers & Models → Preferred Models, pick a specific model for each capability:

Slot	Used for	Typical pick
Chat Model	Default conversation	GPT-4o / Claude 3.5 Sonnet / Gemini 2.5 Pro
Embedding Model	Memory system retrieval vectors	text-embedding-3-large
TTS Model	Read AI replies aloud	OpenAI tts-1
Speech-to-Text	Voice input	whisper-1
Image Generation	Configured separately in Image Generation	gpt-image-1 / Gemini Imagen / DALL·E
Daily Pulse Model	Daily Pulse generation	Pick something cost-effective

These are all independent — you can run chat on GPT-4o, STT on Whisper, and Daily Pulse on Claude.

Multimodal: Images, Voice, Files

"Multimodal" means the AI handles more than text. Prerequisite: the model you choose must support that modality.

Image Input

• From library: chat → + → Choose Photo, multi-select supported
• From camera: + → Take Photo
• Supported: JPG / PNG / HEIC / WebP (auto-transcoded)
• Compatible models: GPT-4o, Claude 3.5/4, Gemini 1.5/2.5

Voice Input (two modes)

Configure under Settings → Speech Input. Once enabled, a mic button appears near the input.

Two delivery modes:

Mode	Behavior	Compatible models
Transcribe first	STT model (Whisper, etc.) converts to text before sending	Any chat model
Send as audio attachment	Audio file goes to the AI for direct listening	GPT-4o Audio / Gemini / any model with native audio input

File Attachments

+ → Choose File sends any file from the Files app. Note:

• Text-based (TXT, Markdown, JSON, Swift, Python, …) — content gets read out and added to context
• PDF — text extraction is attempted
• Binary — sent as-is with a warning that the model may not parse it

TTS — Read AI Aloud

First time:

Settings → Conversation → Text-to-Speech (TTS) → pick a TTS model

The TTS model must be a TTS-capable model you've added under some provider (tts-1, tts-1-hd, etc.).

Usage:

• Long-press any AI bubble → Read Aloud
• Or auto-play when a reply finishes (toggle in settings)
• A floating control appears during playback (pause / stop / speed)

Does TTS Work Offline?

No — TTS hits the model's API, so it needs the network. Streaming makes it feel as smooth as system TTS.

If you want pure offline, the Extended Features section has a toggle to use iOS's built-in AVSpeechSynthesizer instead, but the voice quality is lower.

Reading Experience: Long AI Output as a Readable Document

ETOS's rendering layer is more than vanilla Markdown. By default:

Markdown Enhancements (Settings → Display & Experience → Backgrounds & Visuals)

Toggle	Default	Effect
Enable Markdown	On	Off shows raw text everywhere
Enable Advanced Renderer	On	Code highlighting, math formulas, Mermaid
Auto-Preview Thinking	Off	Reasoning model thinking panel auto-expands
Enable Liquid Glass	On	iOS 26 Liquid Glass effect
Top Blur Fade	On	Soft top edge when scrolling

Code Block Features

• Syntax highlighting for 100+ languages
• Copy button in the corner with feedback animation
• iOS code preview — long-press a Swift/SwiftUI block to launch iOS's native code preview with syntax tree
• Collapse — blocks longer than 50 lines auto-fold

Math / Mermaid

• LaTeX: $inline$ and $$display$$
• Mermaid: ```mermaid fences auto-render

Display System

Custom Fonts

Settings → Display & Experience → Backgrounds & Visuals → Font Fallback:

• Import custom font files (WOFF / WOFF2 / TTF / OTF)
• Slot priority (CJK font / Latin font / monospace)
• Fallback chain (which font to pick when a glyph is missing)

Use cases:

• Specific CJK font for mixed CJK-Latin text
• Screen-optimized fonts like Sarasa Gothic SC for long-form reading

Background and Bubbles

• Enable background image → upload → blur and opacity
• Auto-rotate background per new conversation or on a timer
• Hide assistant bubble — AI replies render directly on the background like a document

Advanced

Preferences (Global Default Parameters)

Settings → Conversation → Preferences controls:

Field	Effect
Global System Prompt	Default system prompt for new sessions. Multiple named profiles supported.
AI Temperature	0–2. Higher = more creative, lower = more deterministic. 0.7 is balanced.
AI Top-P	0–1. Nucleus sampling alternative to temperature. Usually only tune one.
Enable Streaming	Off waits for full generation before showing. Toggle off when debugging a finicky relay.
Enable Response Speed Metrics	Show chars/sec, time-to-first-token per reply
OpenAI Stream Include Usage	Whether OpenAI-compatible streams should carry usage. Disable for relays that omit it.
Enable Auto Session Naming	Auto-name a session after the first reply
Enable Reasoning Summary	Reasoning models generate thinking summaries
Max History Messages	Upper bound on history sent to the model. See below.
Lazy Load Message Count	How many messages to load when opening a session

Tuning "Max History Messages"

A consequential value:

• Low (≤10): saves tokens, AI forgets fast
• High (50+): AI remembers more, requests get expensive
• Recommended: leave around 30 by default, adjust to taste

Global System Prompts

A common pattern:

1. Create multiple profiles: "Default", "Code", "Translation", "Roleplay", …
2. Independent content per profile
3. Switch per conversation in the chat screen or when creating a new session

System Time Injection

In Preferences, Inject System Time auto-includes the current system time in every request. Add Periodic Time Landmark to inject a fresh timestamp every N minutes inside the conversation, keeping the AI grounded in real time during long sessions.

Model-Level Advanced Parameters

In Settings → Providers & Models → [a specific model] → Model Settings you can:

• Override defaults (temperature, top_p, max_tokens) per model
• Write parameter expressions — template syntax to compute parameters dynamically
• Edit the raw JSON body as last-mile customization

Parameter Expression Example

Conceptual, not a copy-paste recipe:

max_tokens = clamp(messages.last.content.length * 2, 512, 8192)

Means: scale max_tokens with the last message's length, clamped to [512, 8192].

Raw JSON Body

Some non-standard endpoints want extra fields in the body (e.g. some relays want extra_body.thinking.budget_tokens = 32000 for Claude). This is the last hook. Most users never need it — only when "all standard fields look right but the endpoint rejects."

Session Import/Export

ETOS doesn't lock data in.

Import (Settings → Sync & Backup → Third-Party Import):

• Cherry Studio full backup
• RikkaHub export
• Kelivo export
• ChatGPT official conversations.json
• ETOS bundles (full config)

Export (long-press a message in chat):

• PDF
• Markdown
• TXT

Usage Analytics

Settings → Extended Capabilities → Usage Analytics breaks down token usage and spend by provider / model / date. Requires per-model pricing:

Settings → Providers & Models → [model] → Pricing — fill in prompt / completion rates from the vendor's pricing page. With pricing set, ETOS computes spend automatically.

Next

• Give the AI tools → Tools & MCP
• Let the AI remember things across sessions → Memory & Worldbook
• Get proactive briefings → Daily Pulse