Use case

Best local LLM for data extraction

This page does not name a winner. The title is the job. You test the suggested models on your own prompts, then decide.

The catalog has no extraction tag. This page routes the job as documents and RAG.

Catalog label: documents and RAG. A catalog tag is a routing label, not a benchmark, and this site has no measured score for this job.

Test this use case

Example prompts

• From this note, return JSON with name, amount, and date. Use null when a field is missing. Note: Ada Klein paid 42 euros on 3 March. No invoice number.
• Extract the company, the city, and the deadline from this sentence. If one is absent, say missing. Sentence: Northwind in Antwerp needs the file by Friday.
• List every date and amount in this paragraph, and ignore the rest. Paragraph: A deposit of $200 was paid on 2 May. The balance of $50 is due on 1 June.

The compare page can drop one of these in with one click. They are starters, not a test set we scored.

Open models that carry this tag

Alphabetical. Order here is not a ranking. The compare suggester reorders them from your languages, context, privacy, and budget, and it says why.

• Gemma 3 27BCatalog tags: rag, translation, creative. Routing labels, not benchmarks.
• LFM2.5 2.6BCatalog tags: chat, reasoning, rag. Routing labels, not benchmarks.
• Llama 3.1 8B InstructCatalog tags: chat, code, rag. Routing labels, not benchmarks.
• Llama 3.3 70B InstructCatalog tags: chat, rag, reasoning, code. Routing labels, not benchmarks.
• Mistral Small 3.2 24BCatalog tags: chat, translation, rag, code. Routing labels, not benchmarks.
• Qwen2.5 72B InstructCatalog tags: translation, rag, code, reasoning. Routing labels, not benchmarks.
• Qwen3 32BCatalog tags: code, reasoning, translation, rag. Routing labels, not benchmarks.

The compare also keeps one frontier reference, Claude Sonnet 5.5. It is not open weights, and it is not a hardware recommendation.

Size hardware after you know the model · Hardware catalog · All use cases