Choosing a language model

We're model-agnostic — the right model depends on the task, your budget, and your data requirements. Here's how the options compare and how we decide.

Open vs API models

• Open models (Llama, Mistral, Qwen, DeepSeek, Gemma…) can be self-hosted: your data stays on your servers, costs are predictable at scale, and you avoid per-vendor lock-in.
• API models (GPT, Claude, Gemini…) are fastest to start with and often have an edge on the hardest reasoning and most natural language — but data is processed by the provider and you pay per use.

What to weigh

• Quality on your task — reasoning, writing, or structured output
• Cost — per-token API spend vs hosting your own
• Latency — how fast replies need to be (voice and live chat are sensitive)
• Languages — some open models are notably stronger multilingually
• Context window — how much it can read at once
• Tool / function calling — for agents that take actions
• Data residency — does data need to stay in-house?

Rules of thumb

• Structured, tool-driven tasks (scheduling, order lookups) run fine on mid-size open models
• Grounded Q&A (RAG) depends more on retrieval quality than raw model size
• Creative or high-stakes reasoning is where frontier API models earn their cost
• Sensitive data leans toward self-hosted open models

How we choose

For each use case we shortlist a couple of options, test them on your real examples, and recommend the best balance — then revisit as your volume and needs grow. Every agent's page lists a recommended starting point.