YuanLiuuuuuu

Researcher at WeChat AI, focusing at large multimodal model and large language model.

[YuanLiuuuuuu @a33668874586]

@jerryjliu0 Another benchmark just dropped to test our model's OOD performance — perfect timing.👍

[YuanLiuuuuuu @a33668874586]

@GVoisco39549 Could you please share examples where POINTS-Reader fails, so we can use them to improve the model?

[YuanLiuuuuuu @a33668874586]

@zenstyle @mervenoyann This is the comparison between POINTS-Reader and Docling on OmniDocBench. (lower is better)

[YuanLiuuuuuu @a33668874586]

Many thanks for sharing! 👏👏 POINTS-Reader is a powerful end-to-end document OCR model with high throughput, making it well-suited for large-scale document extraction. The live demo is online, supported by @prithiv_003 , — please give it a try!

https://t.co/nbe57EPXC1

[merve @mervenoyann]

4B OCR with Apache-2.0 license outperforming Mistral OCR 🔥

Tencent released Points-Reader, it's a new model firstly trained on Qwen2.5VL annotations and then self-trained on real data

in many benchmarks, it performs better than Qwen2.5VL and MistralOCR!

[YuanLiuuuuuu @a33668874586]

@NigelHiggs7 @mervenoyann We have not specifically compared POINTS-Reader with it, but we have achieved state-of-the-art performance among existing document OCR models. For more details, please visit our GitHub page.
https://t.co/5d9FFtxtgv

[YuanLiuuuuuu @a33668874586]

@mervenoyann Thank you for sharing, Merve! The performance metrics for POINTS-Reader in the paper are not the latest. We have continued to enhance its performance after our submission to EMNLP, and the latest results, which are significantly higher, are presented in the table below.👇

[YuanLiuuuuuu @a33668874586]

@venki_lfc @mervenoyann For English, you can try our POINTS-Reader, https://t.co/rDtAcvZcnX

[YuanLiuuuuuu @a33668874586]

@casper_hansen_ @mervenoyann The performance is comparable, but the throughput is significantly higher, making it suitable for large-scale extraction scenarios.

[YuanLiuuuuuu @a33668874586]

Introducing our new vision-language model for end-to-end document conversion! 🔥🔥🔥

It delivers state-of-the-art performance on OmniDocBench with blazing-fast throughput.
🚀🚀🚀

Read the paper: https://t.co/zFqHDQLpIZ

[YuanLiuuuuuu @a33668874586]

@KennyUTC 🌶️💩💉🐮🍺

[YuanLiuuuuuu @a33668874586]

@gunsnrosesgirl3 @Grok What's in this video

[YuanLiuuuuuu @a33668874586]

@WenhuChen Look forward !

[YuanLiuuuuuu @a33668874586]

@WenhuChen @jbohnslav What do you think is the key to achieving o1-like performance in the vision-language domain? At this stage, do we need to introduce RL?

[YuanLiuuuuuu @a33668874586]

@jbohnslav When will the dataset be released?

[YuanLiuuuuuu @a33668874586]

Happy to see that POINTS1.5 ranks first on this leaderboard among models with fewer than 10 billion parameters🚀🚀🚀

Paper: https://t.co/1IziahPLTR
Github: https://t.co/k7qE3rgkUi
Model: https://t.co/0xOxdx6xFS

[KennyUTC @KennyUTC]

OpenCompass has established a leaderboard to evaluate complex reasoning capability of LMMs, consisting of four advanced multi-modal math reasoning benchmarks. Currently, Gemini-2.0-Flash took the 1st place. DM me to suggest more benchmarks and models to this LB.

[YuanLiuuuuuu @a33668874586]

Many thanks to @WenhuChen for sharing our work.

[Wenhu Chen @WenhuChen]

Great work to build VLMs that are more geared towards real-world application. Also glad to see that they evaluated on our MEGA-Bench.
https://t.co/NFh0RHboCn

[YuanLiuuuuuu @a33668874586]

Thanks for sharing. We will continue to develop our WePOINTS series of large multimodal models, bringing more high-quality work to the community.

Paper: https://t.co/VSXQEPrvs2
Github: https://t.co/k7qE3rgkUi

[AK @_akhaliq]

POINTS1.5

Building a Vision-Language Model towards Real World Applications

[YuanLiuuuuuu @a33668874586]

@AdinaYakup We will regularly release our work in the future, and stay tuned💪.

[YuanLiuuuuuu @a33668874586]

Many Thanks to Yakup for sharing💗

[Adina Yakup @AdinaYakup]

POINTS1.5🔥 new vision language model released by Wechat AI @Weixin_WeChat

Model: https://t.co/iZ7INhngME
Paper: https://t.co/asrGPdIurB
✨ Dynamic resolution with NaViT encoder, no tiling required
✨ English & Chinese
✨ Smarter visual tuning with rigorous dataset filtering
✨ Trained on <4B tokens, yet POINTS1.5-7B leads OpenCompass (<10B params)

[YuanLiuuuuuu @a33668874586]

Many Thanks to Tiezhen for sharing💗

[Tiezhen WANG @Xianbao_QIAN]

A new model from @Weixin_WeChat The largest all-in-one messenger app from @TencentGlobal

- VL, ranking top on OpenCompass Vision multimodal leaderboard for &lt;10B models
- 8B, fine-tuned on @Alibaba_Qwen 2.5 7B chat model
- Dynamic high resolution with naViT-style vision encoder