AI model picker: Claude, Groq and local engines compared
The model picker chooses which engine answers you and states the three facts that decide it: what it costs, where your words go, and what the model does badly. Claude, Groq and any locally installed model appear in one list.
Senzoukria · Documentation · Updated September 2026
Where to find it
- Where
- AI page → model button in the chat header; defaults also in Settings → AI
- Default engine
- Claude, with your own API key
- Default Groq model
- llama-3.3-70b-versatile
- Reads screenshots
- Claude only; Groq and local engines receive text alone
- Runs app tools
- All three engines
What it does
The picker lists Claude, Groq and every model the local engine reports as actually installed. No local model is hard-coded: the list is asked of the engine, which is the only thing that knows what has been downloaded, so a model you do not have never appears as available.
Each row carries three columns that usually go unsaid — cost, where your words go, and what the model is weak at. Claude is paid per message on your key and sent to Anthropic. Groq is free on your own key, fast, open models, but rate-limited per minute and weaker on long reasoning and code. A local model never leaves the machine and costs nothing, and the row says plainly that an 8B model is below Claude on code and long reasoning.
Vocabulary is separated on screen for a reason: Ollama is the engine, the program listening on a local port; qwen3:8b is the model, the weights it loads. The engine has its own line.
Settings
| Setting | Default | What it changes |
|---|---|---|
| Default engine (Settings → AI) | Claude | Claude, Groq or Local engine — which provider answers by default |
| Groq model | llama-3.3-70b-versatile | The exact Groq model id; the list is read from the Groq API on your key |
| Context window | 8k | Tokens the local engine loads the model with: 4k, 8k, 16k, 32k or 64k. Bigger needs more VRAM and makes Ollama reload the model |
| Web access for the assistant | On | Search the web and read pages — Anthropic's search for Claude, DuckDuckGo through the app for Groq and local models |
| Tools (read the chart, run scripts) | On | Lets the agent read chart, account, calendar, news, backtests, scripts and autopilot — read-only |
| Fall back to the local model | Off | Used only when a Claude request fails; the reply then carries the badge of the engine that wrote it |
How to use it
- Pick Claude when the task is writing indicator code or following a long chain of reasoning, and when you want screenshots to be read.
- Pick Groq to start without paying: a free key from the Groq console, open models, tool calls and web search — at the cost of per-minute rate limits and no image reading.
- Pick a local model when nothing may leave the machine. The picker shows the exact model name, never the word "Claude", so a reply written by an 8B model is never mistaken for a Claude one.
- If a local model you selected has since been deleted from the machine, the picker marks the choice as unavailable instead of letting a send fail silently.
Limits and pitfalls
- "Local" with no model selected does not silently fall back to Claude. The request is refused and says so, because falling back would send into the cloud what you meant to keep on the machine.
- Embedding models turn text into vectors and cannot answer; the picker flags them rather than letting one be chosen as the assistant.
- Groq's free tier is rate-limited per minute, and the context window setting applies to local engines only — Claude and Groq manage their own.
- Whatever engine you choose, its cost is billed by that provider on your key. Market data remains billed separately by your data provider.
Related pages
This page in other languages
Frequently asked questions
- Is Ollama a model?
- No. Ollama is the engine — the program on your machine listening on a local port. qwen3:8b is a model it loads. The picker keeps them on separate lines for exactly this reason.
- Why can't I attach a screenshot on this engine?
- The local path sends plain strings to the engine and Groq's rows are text-only, so neither receives images. The picker states it before you attach, rather than letting the send fail afterwards.
- What does the context window setting do?
- It tells the local engine how many tokens to load the model with. 8k fits the assistant's prompt plus a short conversation; larger values need more VRAM and cause Ollama to reload the model.