Model Management

Wilson supports 10 LLM providers with dozens of models. Use /model to switch between them and /pull to download local models.

Switching Models

`/model` — Interactive Selector

Run /model without arguments to open the interactive model selector:

Pick a provider (Ollama, Transformers.js, OpenAI, Anthropic, etc.)
Pick a model from the provider’s list, or type a custom model name
Wilson switches immediately — the next message uses the new model

/model

`/model <name>` — Direct Switch

Skip the selector by passing a model name directly:

/model ollama:qwen3:8b
/model gpt-4.1
/model claude-sonnet-4-6
/model transformers:HuggingFaceTB/SmolLM3-3B-ONNX

Downloading Models

`/pull <model>` — Download for Local Use

Download Ollama or Transformers.js models with progress reporting:

/pull qwen3:8b                                        # Ollama model
/pull transformers:HuggingFaceTB/SmolLM3-3B-ONNX      # Transformers.js model

For Ollama models, this calls ollama pull under the hood. For Transformers.js models, it downloads ONNX weights from HuggingFace Hub to ~/.openaccountant/models/.

Recommended Ollama Models

Tool-Calling Models

These models work with Wilson’s agent tool system — they can call tools like search_transactions, categorize, and import_csv.

Model	Size	Description
`qwen3:0.6b`	523 MB	Alibaba — smallest tool-calling model
`qwen3:4b`	2.5 GB	Alibaba — rivals 72B quality, tool-calling
`qwen3:8b`	5.2 GB	Alibaba — best balance, tool-calling
`granite4:3b`	2.1 GB	IBM Granite 4 — tool-calling, 128K context
`granite4:tiny-h`	4.2 GB	IBM Granite 4 MoE — 7B, tool-calling
`ministral-3:3b`	2.1 GB	Mistral — vision + tool-calling, 256K context

General-Purpose Models

Good for conversation and analysis, but may not reliably call tools.

Model	Size	Description
`gemma3:4b`	3.3 GB	Google — multimodal, 140+ languages
`smollm2:1.7b`	1.8 GB	HuggingFace — tiny but capable
`gemma3:1b`	815 MB	Google — ultralight, 32K context
`granite4:350m`	708 MB	IBM — smallest Granite, edge-ready

Default Model

Set DEFAULT_MODEL in .env for headless mode and cron jobs:

DEFAULT_MODEL=ollama:qwen3:8b

This is used when Wilson starts without an interactive model selection — for example, with wilson --run "..." or wilson --sync.

Model Prefixes

The model name prefix determines which provider handles the request:

Prefix	Provider	Example
`ollama:`	Ollama	`ollama:qwen3:8b`
`transformers:`	Transformers.js	`transformers:HuggingFaceTB/SmolLM3-3B-ONNX`
`openrouter:`	OpenRouter	`openrouter:anthropic/claude-3.5-sonnet`
`litellm:`	LiteLLM	`litellm:gpt-4o`
`claude-`	Anthropic	`claude-sonnet-4-6`
`gpt-`	OpenAI	`gpt-4.1`
`gemini-`	Google	`gemini-3-flash-preview`
`grok-`	xAI	`grok-4-0709`
`kimi-`	Moonshot	`kimi-k2-5`
`deepseek-`	DeepSeek	`deepseek-chat`

No prefix defaults to OpenAI.