Author(s): Chew Loong Nian – AI ENGINEER

Originally published on Towards AI.

Gemini 3.6 Flash Reads Charts 14 Points Worse Than the Model It Replaced — LlamaIndex Ran the Numbers

Yesterday everyone cheered Gemini 3.6 Flash for topping the computer-use leaderboards. Then Jerry Liu, the CEO of LlamaIndex, ran it through a document-understanding benchmark and found the opposite story hiding underneath: on charts, the shiny new Flash model scores 14 points worse than the model it replaced — 45.0 down to 31.0. If you use Gemini Flash to read documents, the “upgrade” is a downgrade.

The article explains that ParseBench, which focuses on real document-parsing capabilities (tables, chart understanding, layout fidelity, and text faithfulness), shows Gemini 3.6 Flash regressed compared with Gemini 3.5 Flash: chart-rule pass rate dropped 14 points (45.0→31.0) and the overall document score fell (69.9→66.8), even though coding/agentic performance improved. It attributes the change to post-training tradeoffs—capacity gained for coding and reasoning appears to come at the expense of visual document recognition. A “Flash Lite” comparison reinforces the same pattern of skill tradeoffs, and the piece urges readers not to rely on leaderboards or auto-upgrade to “latest” for document pipelines. Instead, it recommends pinning model versions, running side-by-side evaluations on representative pages from your own documents, using schemas and deterministic validation (e.g., for invoices) to make probabilistic OCR safer, and choosing the version that wins for your specific workload (chart/table-heavy vs layout-focused vs agentic tasks).

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI