Skip to contents

Default VLM model per provider

Source: R/vlm-providers.R
default_model.Rd

Returns a sensible default model name for the requested provider. These defaults are picked for **vision capability** (multimodal) AND **structured-extraction reliability** – the two things the soilKey extraction layer needs.

Usage

default_model(name)

Arguments

name

Provider name; one of "anthropic", "openai", "google", "ollama".

Value

Character scalar with the default model identifier.

Details

Defaults (as of v0.9.11):

  • anthropic = "claude-sonnet-4-7" – the strongest Claude vision model at our 2026 cutoff for document / photo extraction.

  • openai = "gpt-4o" – text + vision.

  • google = "gemini-2.0-pro" – successor to 1.5 with longer context + better multimodal grounding.

  • ollama = "gemma4:e4b" – Gemma 4 edge multimodal (text + image; audio also). For larger contexts use "gemma4:31b"; for cloud-only offload via Ollama, "gemma4-cloud:31b". Pull the desired size first with ollama pull gemma4:e4b.

Users can override at any time:


vlm_provider("ollama", model = "gemma4:31b")
vlm_provider("ollama", model = "gemma3:27b")  # back-compat
vlm_provider("ollama", model = "qwen2.5vl:32b")  # any pulled model