Choosing an AI Model

AI is entirely optional in Finzytrack — every core feature works without it. If you do choose to enable it, the model you pick matters a lot. This page covers what AI unlocks, what kind of model is capable enough to be useful, and which models and providers we have tested.

For the YAML schema and field-by-field reference of the ai config section, see LLM Configuration. For what data leaves your machine in each AI feature, see Data Shared with AI.

What AI unlocks

Statement parsing — Upload a CSV, PDF, or image and let the AI extract transactions without writing rules.
Import rule creation — Ask the assistant to generate CSV/XLS import rules from a sample file.
Natural-language entry — Describe transactions in plain English.
Auto-categorization — Let the AI suggest expense/income accounts for imported transactions.
Financial queries — Ask questions about your data in natural language.

Model requirements

Finzytrack’s AI features require a capable model. Specifically, the model needs to support:

Tool calling (function calling)
Reasoning and coding ability
A context window of 128k tokens or more
32B or more active parameters

Most smaller models that run locally on typical consumer hardware are unlikely to produce reliable results. The AI features involve complex multi-step tasks like parsing financial statements, creating structured rules, and executing query plans — these require a model with strong instruction-following and tool-use capabilities.

Tested models

The table below lists specific models we have used with Finzytrack and found to work acceptably across the full range of AI features (statement parsing, rule creation, natural-language entry, and the assistant). It grows as we test more. Even the models below have occasional hiccups on specific tasks — a malformed rule, a query that needs a second try, a parse that misses a column — so treat the assistant’s output as a draft and review it before committing.

Model	Reasoning?	Notes
GLM-5	Yes	Open-source model by Zhipu AI. Works satisfyingly well across all AI features.
GLM-4.7	Yes	Earlier open-source model by Zhipu AI. Also works well across all AI features; GLM-5 is preferred where available.
Qwen3-Coder-480B-A35B-Instruct	No	Open-source non-reasoning instruct model by Alibaba. Strong tool calling and structured output, and avoids the reasoning-loop failure mode entirely. A good choice if you’ve hit reasoning hangs with the GLM models.

Models marked Reasoning: yes can occasionally hang or return empty answers when the model gets stuck thinking; see Reasoning Models for the failure mode and what to do.

A note on providers

We only test against providers with strong privacy commitments, since some of your financial data may be included in prompts. So far that has meant Synthetic, which runs open-source models in private datacenters and states it never trains on user data and never stores API prompts or completions. We are not affiliated with Synthetic — just happy users.

The specific models above are not tied to Synthetic — they are open-source and may be offered by other providers, and providers routinely sunset specific versions (e.g., a provider might drop GLM-5 in favor of GLM-5.1). Pick the most recent version of a listed family that your provider currently offers. If you choose a provider we haven’t mentioned, review its privacy policy carefully; see Data Shared with AI for what is sent in each feature.

In the future, Finzytrack AI will offer a cloud-hosted model that requires no configuration on your part, preserves your privacy, and is well-suited for financial imports and analysis.

Trying AI before committing to a provider

If you want to explore Finzytrack’s AI features before signing up with a paid provider, OpenRouter offers a free tier with no credit card required. Sign up, generate an API key, and browse their free models — look for one that meets the requirements above (tool calling, 128k+ context, 32B+ parameters). Set the API URL to https://openrouter.ai/api/v1 in Settings > AI.

Be aware that privacy policies vary across free models on OpenRouter — some providers may use your prompts for training. Check the model card for the specific model you choose before sending any sensitive financial data. For regular use, we recommend switching to a provider with explicit privacy commitments, such as Synthetic.

Provider modes

You configure AI during the setup wizard or later from Settings > AI. There are two provider modes:

OpenAI-compatible — Works with cloud providers (Synthetic, OpenAI, Groq, etc.) and local servers (LM Studio, Ollama). Enter the API URL, API key (if required), and model name.
Anthropic — Direct connection to Anthropic’s API. Enter your API key and model name.

See LLM Configuration for the full reference of AI settings.