If you are evaluating OCR-capable multimodal models for production work, the question is not just whether a model can read text in an image. The real question is whether it can do it consistently, with structured outputs, across the kinds of inputs teams actually send in production: screenshots, receipts, tables, rotated documents, scene text, and low-resolution UI captures.
We ran a live benchmark for youtu-vita through our OpenAI-compatible API path and scored it on a controlled OCR test set. This article shares the actual test data, what the model passed, where it struggled, and what kind of OCR workloads it looks good at right now.
This benchmark was run on June 24, 2026.
We used a local generated OCR benchmark set with these eight cases:
document_basicreceipt_totalui_settingstable_statementscene_text_signboardrotated_documentlow_res_small_textchart_with_legendThe benchmark images were generated locally so the test would be stable and reproducible, rather than depending on external image hosts. Each request used the same OpenAI-compatible image input shape and the same JSON-only output instruction. The model was asked to:
The benchmark files used for this run:
.tmp/ocr_compare_manifest.generated.json.tmp/ocr_benchmark_assets/.tmp/ocr_model_compare_20260624_153400.jsonEach case was scored on four dimensions:
The final score per case is a weighted total. A score of 1.000 means the model fully passed that case under this benchmark.
The full 8-case run completed successfully.
100%0.8753794 ms4088 ms15284 ms| Case ID | Category | Status | Score | Latency |
|---|---|---|---|---|
document_basic | Document OCR | 200 | 1.000 | 3852 ms |
receipt_total | Receipt OCR | 200 | 1.000 | 3616 ms |
ui_settings | UI screenshot OCR | 200 | 1.000 | 3794 ms |
table_statement | Table OCR | 200 | 1.000 | 3434 ms |
scene_text_signboard | Scene text | 200 | 1.000 | 1661 ms |
rotated_document | Rotated document | 200 | 1.000 | 2908 ms |
low_res_small_text | Small text / low resolution | 200 | 1.000 | 4088 ms |
chart_with_legend | Chart reasoning | 200 | 0.000 | 15284 ms |
For this run, youtu-vita was strong on the OCR tasks most teams care about first:
It correctly extracted:
On document_basic, it returned a full structured transcription and correctly answered:
title = Quarterly Operations Summarydate = 2026-06-24document_id = AB-123456It handled receipt-style layout correctly on receipt_total, including:
That matters because receipt OCR often breaks on spacing, alignment, or repeated numeric fields. In this run, youtu-vita passed the receipt case with a full score.
On ui_settings, it correctly captured:
It also returned the structured answers we asked for:
primary_cta = Continueerror_code = E102That makes it promising for support automation, QA workflows, screen parsing, and screenshot-based extraction tasks.
On table_statement, the model passed the table case with a full 1.000 score in this run.
That is important because table OCR is often where vision models look good at plain text but fail at row-column alignment. In this benchmark, youtu-vita handled the table extraction cleanly enough to pass both the visible text requirements and the structured answer checks.
On rotated_document, the model also scored 1.000.
That suggests it is not limited to perfectly upright scanned pages. If your OCR workflow includes phone photos, skewed uploads, or documents captured in the wild, this is a meaningful result.
One of the most practically useful passes in this run was low_res_small_text, which also scored 1.000.
That case is closer to real dashboard and UI OCR than a clean printed PDF. If you need to read release notes, settings screens, logs, or admin panels from screenshots, this is a positive signal.
The weak spot in this run was not standard OCR. It was chart reasoning.
On chart_with_legend, youtu-vita returned HTTP 200 but scored 0.000. It also took much longer than the rest of the test set at 15284 ms.
That tells us two things:
This matters because many teams group all “image understanding” into one bucket. That is too coarse. A model can be strong at:
and still be weak at:
Based on this live run, youtu-vita looks strongest for these workloads:
If your workload is mostly “read the text, extract the fields, give me clean JSON,” this benchmark suggests youtu-vita is already useful.
If your workload is “understand a chart, infer trends, compare the latest month, and reason visually,” this benchmark does not support calling it strong there yet.
Many OCR evaluations are too soft. They say a model is “good at image understanding” after a single logo or document test. That does not help if you need to decide whether to route:
into a production OCR pipeline.
This benchmark is more useful because it separates:
For this run, youtu-vita was clearly stable across the OCR-heavy categories.
For this live benchmark, youtu-vita was the most stable OCR model we tested in the current environment.
The actual results were:
100% request success across the full 8-case OCR set0.875 average total score1.000 on 7 of 8 OCR-oriented casesThat makes youtu-vita a strong candidate if your main need is text extraction from images, especially for:
It does not mean it is the best choice for every kind of vision workload. But if your problem is OCR, not visual analytics, this is one of the clearest positive live results we have seen in this environment.
If you want to run the same OCR benchmark shape yourself, we used:
The local result summary used for this article:
If you want a follow-up article comparing youtu-vita directly against Gemini or GPT-family vision models on the same OCR set, that should be written as a stability + latency + category-by-category comparison, not just a single overall score.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。