Llama 3.1 8B Instruct and Qwen3-8B are both small enough to consider for local text workloads. Compare the checkpoint, runtime, and output format together. A model that produces a convincing conversational answer may still make costly errors when the output must fit a fixed schema.
Check context and terms first
Meta documents a 128K context for Llama 3.1 and distributes it under its community license. Qwen documents 32,768 native context tokens for Qwen3-8B and up to 131,072 with YaRN. Its model card lists Apache 2.0. Check the card and license of the specific artifact you download, particularly if somebody has repackaged or quantized it.
Qwen also documents thinking and non-thinking behavior for this checkpoint. For a short extraction task, record which mode you used. Reasoning text, longer answers, and chat-template differences can otherwise make a timing comparison hard to interpret.
Use a test with exact answers
Make ten fictional invoices as plain text. Ask for invoice ID, issue date, currency, subtotal, and a null value when a field is missing. Include one invoice with two dates and one that quotes a payment amount from a previous invoice. Run both models with a valid template and the same output rules.
Score valid JSON separately from correct extracted values. Measure response time on the same machine, quantization class, and input sizes. If both models meet the accuracy threshold, decide using the actual license and latency requirements of your deployment.
Continue with Oakgen Agent
Open in Oakgen Agent
Review the prefilled prompt before sending.


