The LLM is the one part of Yumii that is (by default) a cloud call — her ears and voice run on your machine, but the thinking uses a provider you choose. Pick it in Dashboard → Settings → Mind (LLM), paste the matching key, save, and restart.
The four providers
| Provider | Default model | Why pick it |
|---|---|---|
| Ollama (default) | minimax-m3 | Free tier with a huge context window (1M) — tool schemas and long conversations fit without squeezing. Cloud with one key, or point it at a local Ollama |
| Groq | qwen/qwen3.6-27b | Extremely fast inference (matters for voice), free tier, no card |
| OpenAI | gpt-4o | Strongest general capability |
| Anthropic | claude-3-5-sonnet | Most nuanced conversational quality |
For a voice companion, latency is quality — and so is context room. Ollama's default balances both; Groq is the speed play when your conversations are light on tools.
Choosing the exact model
Two providers take a free-text model name in the dashboard:
- Ollama model — any model on your Ollama account (
minimax-m3,gpt-oss:120b, …). Note: some models need a paid Ollama subscription and will refuse on the free tier. - Groq model — any Groq-hosted model. Each model has its own free-tier quota bucket, so switching models also switches quotas.
Ollama: cloud or fully local
- Ollama Cloud (default): create a key at ollama.com, paste it, done.
- Local Ollama: run Ollama on your own hardware, set
OLLAMA_BASE_URLin~/.yumii/config.jsonto your server (e.g.http://127.0.0.1:11434), leave the key empty. This is the path to a fully offline Yumii — voice, ears, and mind all local — if your machine can carry a capable model.
Free tiers: what to expect
- Quotas. Ollama's free tier is a weekly allowance; Groq's is daily. A heavy day (or a testing spree) can exhaust them — she says so with a clear error card, and you can switch providers in the dashboard or wait for the reset.
- Request-size caps (Groq). Groq's free tier caps a single request at roughly 8–12k tokens. Yumii budgets for it — compact tool loading, truncated tool results, a bounded history window — but connecting several toolkits at once is what breaks the budget first (see Tools). Ollama's 1M context makes this a non-issue there.
Behind the scenes
- Prompt caching. Context is ordered so the static parts form a stable prefix providers can cache — faster first tokens, every turn.
- Tool-call resilience. Schemas are sanitized at bind time and malformed calls retried once, so a model hiccup degrades to a spoken apology, not a crashed turn.
- A second, cheaper call. Memory review runs in the background on a small fast model from your same provider — part of the same quota, not a separate account.