Open Source AI Models for Accounting Firms

Open source AI gives you permission to use, inspect, change and share an AI system. Downloadable model weights alone do not establish those freedoms.

For an accounting practice, the licence decides permitted use. The way you run the model decides who can access client information. These are separate decisions, even when the download is free.

What Open Source AI Means

The Open Source Initiative’s definition requires access to model parameters, training and running code, and detailed information about training data. It also requires permission to use and modify the system for any purpose.

The Open Source AI Definition does not require every training record to be publicly downloadable. It requires enough data information for a skilled person to build a substantially equivalent system.

Model weights are the learned numerical settings that produce a model’s responses. An open-weight release lets you download those settings. The release can still have restrictions on commercial use or provide too little information to meet the full definition.

That distinction changes what an IT provider is promising. A downloadable model with a permissive licence can be useful without establishing that its entire development meets the Open Source Initiative’s definition.

For example, imagine your practice wants to summarise internal procedure notes. You can run downloaded weights inside an approved environment. You still need software around the model and permission to use each component for that job.

Open source also does not mean that a public chatbot keeps your prompts private. A hosted service can use an open model while applying its own terms to the information you enter.

The Best Open Source AI Models

For a small text-only starting point, Qwen3-8B is a practical choice. Mistral Small 3.2 adds image understanding, while DeepSeek-R1 suits reasoning-focused work. Choose the capability your task needs before considering a larger model.

These are specific downloadable releases with commercial-use licences. Their published capabilities support the uses below. A model’s general reasoning ability does not establish that it can apply Australian tax law correctly.

Model release	Licence	Useful accounting task	Limitation that affects the choice
Qwen3-8B	Apache 2.0	Drafting internal explanations from supplied text	Text model; scanned pages need a separate extraction step
Mistral Small 3.2 24B Instruct	Apache 2.0	Reading text alongside document images	Image interpretation still needs checking against the source
DeepSeek-R1	MIT	Reasoning through a written explanation or calculation approach	A reasoning response is not proof of a correct tax treatment

Qwen’s model card documents thinking and non-thinking modes in Qwen3-8B. Use the ordinary response mode for straightforward drafting. A longer reasoning response adds little when the task is simply rewriting a procedure in plain English.

Mistral’s Small 3.2 model card retains Small 3.1’s text and vision capabilities, with changes to instruction following and repetitive responses. It is the more useful option here when the document’s visual layout contributes to the question.

DeepSeek’s R1 repository permits commercial use under MIT terms. Its smaller distilled releases use different underlying models. The repository identifies separate licence origins for the Qwen-derived and Llama-derived versions, so the family name alone is insufficient.

Keep the complete release name in your approval record. For Mistral, that includes the Small 3.2 24B Instruct identifier ending in 2506. A model family’s next release can carry different terms.

Running a local LLM is a separate implementation decision. An open-weight model can also run through a cloud provider, where that provider’s access and retention terms apply.

Open Source AI for Document Processing

Open source document tools turn scanned pages and PDFs into text or structured records that another system can use. Optical character recognition, or OCR, reads characters. Document parsing also attempts to preserve relationships such as table rows and headings.

Docling parses PDFs and other document formats, including page layout and table structure. It exports formats such as JSON and Markdown. Its code uses the MIT licence, while individual models have their own licences.

Tesseract is an OCR engine under Apache 2.0. It is useful when you need text from an image. Reading characters does not itself classify an expense, establish a tax entitlement or match a transaction to a client.

In a hypothetical statement, an extraction tool might place a payment under the neighbouring transaction’s description. Every character could look plausible while the resulting row is wrong. Keep the original page beside extracted records during review.

For client documents, review these outcomes before using extracted figures:

• Each amount stays with the correct description and period.
• Debit and credit signs survive extraction.
• Table headings remain attached to their columns.
• Missing pages and unreadable fields remain visible as exceptions.

Keep interpretation separate from extraction. A model can explain the text it receives, but it cannot recover a missing decimal point reliably by sounding confident. Reconcile material figures with the original document before they enter a workpaper.

The OAIC’s guidance on using AI products recommends keeping personal information out of publicly available generative AI tools. Assess who receives documents and prompts before connecting a client folder to any model.

Licences and Commercial Use

Commercial use depends on the exact model and software licences, including any conditions inherited from a base model. Download access, a free demonstration and permission to use a hosted chatbot do not settle that question.

The MIT licence permits commercial use and modification. Copies or substantial portions must retain the copyright and permission notice. The licence also contains a warranty disclaimer.

Apache 2.0 permits commercial use and contains an express patent grant. Redistribution brings conditions, including retaining the licence and relevant notices, marking modified files and handling any supplied NOTICE file.

A practice using a model internally and an IT provider distributing a modified package have different activities to assess. Do not treat a permissive licence as permission to remove notices from software you pass to someone else.

For a release with custom terms, identify any acceptable-use restrictions, commercial thresholds or attribution requirements. Review the model weights separately from the application, OCR engine and any additional downloaded models.

A model licence grants rights over the licensed material. It does not give your practice permission to disclose client information or establish ownership of every generated answer.

Before approving client work, retain the exact release identifier and its licence. Record where the model runs, who can access its inputs and which outputs require source-document review. Approve the named task against that record, rather than approving every use of a model family.