LlamaIndex has released Turbo, a beta extraction tier for applications that need structured information from documents with less delay. It is available on the LlamaParse Platform and is aimed at medium-complexity documents feeding interactive agent workflows.
Turbo skips the separate parsing pass used by other LlamaParse tiers. Instead, it extracts information directly from page images and processes pages in parallel. On LlamaIndex’s ExtractBench benchmark, Turbo recorded a median processing time of 3.7 seconds per page and an F1 score of 0.84, a combined measure of extraction accuracy and completeness.
The company says Turbo was more than four times faster than both its lower-cost Cost Effective tier and its more accurate Agentic tier. Longer documents showed the largest latency advantage: after about 16 pages, parallel processing reduced average time to roughly half a second per page while the other evaluated systems slowed as document length increased. LlamaIndex also says the closest specialized API in accuracy, Extend, took 3.5 times longer per page in its tests.
Those figures come from LlamaIndex’s own benchmark and configuration, so teams should test representative documents before switching. Turbo prioritizes response time rather than maximum performance on every document type, and its beta status means production users should also plan for changes as the tier develops.