# TextVQA

> TextVQA requires reading and reasoning about text in images. These two own-provider reports use the validation split, with metric and scorer left undisclosed in the original model cards.

Canonical page: https://benchlm.ai/benchmarks/textvqa

- Category: [Multimodal & Grounded](/multimodal-grounded)
- Last updated: May 2026 source revisions; checked October 1, 2026

Evaluation results published by Zyphra. Two original Zyphra TextVQA validation reports; metric/scorer not disclosed.

This page retains 2 selected rows across original provider reports about their own models. Comparison-model columns from those reports are excluded. Source model identities remain separate from catalog profiles.

Each cohort keeps its original split, printed score, source, and disclosed setup. The source revision or report publication date is separate from an evaluation date. Unknown settings remain unreported, and the scores do not enter overall or category rankings.

Separate original-provider cohorts are shown below. Undisclosed metrics and different report setups do not support a shared performance ranking or a pooled score.

- [Zyphra ZAYA1-VL-8B model card](https://huggingface.co/Zyphra/ZAYA1-VL-8B/blob/2967258fbdfdf8c9358dc52b3451db24fe45dfcd/README.md)
- [Zyphra Zamba2-VL-7B model card](https://huggingface.co/Zyphra/Zamba2-VL-7B/blob/060812b62a6eb91820274cf2ec7abc83297534fe/README.md)
- [Original benchmark task](https://arxiv.org/abs/1904.08920)

## About TextVQA

- Year: 2019
- Format: Provider-reported validation scores; metric/scorer not disclosed
- Paper: [Towards VQA Models That Can Read](https://arxiv.org/abs/1904.08920)

The source calls the split val. ZAYA1-VL-8B and Zamba2-VL-7B retain separate model-card cohorts. The named VLMEvalKit-based harness does not establish an exact scorer revision or run setup. Scores stay display only, without a percentage suffix or an inferred accuracy definition.

TextVQA is currently displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.

## Leaderboard (2 own-provider reports)

### Zyphra: ZAYA1-VL-8B model card

Zyphra harness based on VLMEvalKit; metric/scorer and run settings not disclosed.

Source repository modified: 2026-05-17; this is not an evaluation date.

| Source model | Reported score | Metric/scorer | Split | Original provider source |
|---|---|---|---|---|
| [ZAYA1-VL-8B](https://huggingface.co/Zyphra/ZAYA1-VL-8B/blob/2967258fbdfdf8c9358dc52b3451db24fe45dfcd/README.md) | 74.4 | Metric/scorer not disclosed | validation | [Zyphra report](https://huggingface.co/Zyphra/ZAYA1-VL-8B/blob/2967258fbdfdf8c9358dc52b3451db24fe45dfcd/README.md); Pinned official model-card Performance table; own-model column only |

### Zyphra: Zamba2-VL-7B model card

Zyphra harness based on VLMEvalKit; metric/scorer and run settings not disclosed.

Source repository modified: 2026-05-24; this is not an evaluation date.

| Source model | Reported score | Metric/scorer | Split | Original provider source |
|---|---|---|---|---|
| [Zamba2-VL-7B](https://huggingface.co/Zyphra/Zamba2-VL-7B/blob/060812b62a6eb91820274cf2ec7abc83297534fe/README.md) | 81.0 | Metric/scorer not disclosed | validation | [Zyphra report](https://huggingface.co/Zyphra/Zamba2-VL-7B/blob/060812b62a6eb91820274cf2ec7abc83297534fe/README.md); Pinned official model-card Performance table; own-model column only |

Only each provider's own model rows are included. Unknown metric/scorer means the printed value is retained without interpreting it as accuracy, ANLS, or a percentage. Cohorts do not establish a shared run or protocol. No pooled score or cross-cohort rank is calculated.

## FAQ

### What does TextVQA measure?

TextVQA requires reading and reasoning about text in images. These two own-provider reports use the validation split, with metric and scorer left undisclosed in the original model cards.

### Can these source results be compared directly?

Separate original-provider cohorts are shown below. Undisclosed metrics and different report setups do not support a shared performance ranking or a pooled score. Results stay outside weighted overall and category scoring.

### How many source reports are shown?

This snapshot shows 2 selected source reports. Each row retains its original source and disclosed setup.

### Does TextVQA affect BenchLM's overall score?

Not directly. TextVQA is still displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.