Tony Wu ✓

@tonywu_71

Multimodal, RAG, Agents | ColPali co-first author | @centralesupelec 🇫🇷 x @Cambridge_Uni 🇬🇧 | Core Researcher at @hcompany_ai 🧑🏻‍💻

1 565Followers

544Following

572Posts total

29.3KViews on collected posts

Ultimi post

Tony Wu ✓@tonywu_713 Sep

@tonywu_71 congrats Tony! very interesting work and the performance seems to be strong! And i like a lot you put a lot of efforts on the writing and making beautiful charts in the report! 242 views · 9 likes · 0 reposts · 1 replies Open on X →

Tony Wu ✓@tonywu_713 Sep

And many thanks to Raushan Turganbay, @mervenoyann and @art_zucker for helping me land the day-0 transformers integration! And huge kudos to @tomaarsen, the goat maintainer of Sentence Transformers who recently added late-interaction support! 🤗 (15/N) 232 views · 9 likes · 0 reposts · 0 replies Open on X →

Tony Wu ✓@tonywu_713 Sep

Cc to the late-interaction gang 🙏🏼 @lateinteraction @ManuelFaysse @bclavie @antoine_chaffin @raphaelsrty @AmelieTabatta @MaceQuent1 @pteiletche @antoine_edy @bo_wangbo @jobergum @doesdatmaksense @helloiamleonie @CShorten30 (14/N) 376 views · 13 likes · 0 reposts · 1 replies Open on X →

Tony Wu ✓@tonywu_713 Sep

NeoMME began as a side quest between two good friends (thanks @Aurelien_L_ for bearing with me 🫶🏼). We both worked with limited time, but I hope you'll find our work interesting! And thanks to @hcompany_ai for supporting this work and providing the training compute. (13/N) 244 views · 12 likes · 0 reposts · 1 replies Open on X →

Tony Wu ✓@tonywu_713 Sep

We wrote a longer blog post and a 54-page tech report to share all our learnings with the community. You can find these in our 🤗 Hf Collection, which includes the Hf Space demo and the open-weight Apache 2.0 checkpoints. (12/N) https://t.co/1968WUaj3l 274 views · 17 likes · 0 reposts · 1 replies Open on X →

Tony Wu ✓@tonywu_713 Sep

You can use NeoMME-Retriever for visual RAG: given a query, it retrieves the top-k most relevant page images. A VLM then answers the original question using those pages as context. I created an Hf Space so you can try it yourself! (11/N) https://t.co/8pARXLHT1s

0:19

271 views · 13 likes · 0 reposts · 1 replies Open on X →

Tony Wu ✓@tonywu_713 Sep

NeoMME-Retriever is available in 🤗 Transformers. The snippet below scores two queries against two dummy page images using both dense and late-interaction heads. (10/N) https://t.co/x2dKbZjM8X

286 views · 15 likes · 0 reposts · 1 replies Open on X →

Tony Wu ✓@tonywu_713 Sep

For NeoMME-Retriever-260M on ViDoRe v3, high-resolution late-interaction embeddings average about 1.5 MB per page. Token pooling and asymmetric quantization reduce them to 6 kB, 255x smaller, while retaining more than 95% of baseline nDCG@10. (9/N) https://t.co/6c1Z164NoL

1.2K views · 18 likes · 1 reposts · 1 replies Open on X →

Tony Wu ✓@tonywu_713 Sep

With matched 2048x2048 inputs on one NVIDIA L40S, NeoMME-Retriever-260M encodes about 51 pages per second, nearly 2x ColModernVBERT's 26 pages per second. Faster document encoding means less GPU time, so cheaper indexing! ⚡️ (8/N) https://t.co/F6zOncFPiK

336 views · 15 likes · 0 reposts · 1 replies Open on X →

Tony Wu ✓@tonywu_713 Sep

On ViDoRe v3 and using the late-interaction embeddings, NeoMME-Retriever-260M reaches 0.523 nDCG@10, the highest score among evaluated models below 800M parameters. The 800M model reaches 0.556. Both are on the benchmark's model-size Pareto frontier. (7/N) https://t.co/RnpQVmJwII

1.3K views · 16 likes · 1 reposts · 1 replies Open on X →

Tony Wu ✓@tonywu_713 Sep

To get a meaningful downstream evaluation of NeoMME, we fine-tuned the model for visual document retrieval 🔎 We added two jointly trained heads so the model can output both dense and late-interaction embeddings in a single forward pass. (6/N) https://t.co/p3ZAiwPogj

389 views · 12 likes · 0 reposts · 1 replies Open on X →

Tony Wu ✓@tonywu_713 Sep

We pretrain NeoMME from scratch with masked discrete diffusion. Image patches stay visible while the model reconstructs masked text. At high masking rates, the model is forced to use the image instead of guessing from the visible text. (5/N) https://t.co/jIuHrvbTmy

433 views · 15 likes · 0 reposts · 2 replies Open on X →

Tony Wu ✓@tonywu_713 Sep

NeoMME splits images into 32x32 patches and keeps their aspect ratio and resolution. Most layers use sliding-window attention, while every sixth layer and the final layer use global attention. A 16k-token context fits up to two 4K UHD images in one forward pass. (4/N) https://t.c

485 views · 15 likes · 0 reposts · 1 replies Open on X →

Tony Wu ✓@tonywu_713 Sep

With NeoMME, we decided to build a single bidirectional Transformer encoder that processes both text tokens and raw image patches and could be trained entirely from scratch. (3/N) https://t.co/DN7H5UWSZO

776 views · 17 likes · 0 reposts · 1 replies Open on X →

Tony Wu ✓@tonywu_713 Sep

Multimodal encoders like ColPali generally reuse pretrained generative VLMs architecture (vision encoder → projector → LLM). Such models are powerful, but they are usually large and the asymmetry in how modalities are handled make it suboptimal for deployment, fine-tuning, 927 views · 18 likes · 0 reposts · 1 replies Open on X →

Tony Wu ✓@tonywu_713 Sep

👀 Meet NeoMME: a family of 260M and 800M Multimodal-Native Multilingual efficient Encoders One bidirectional Transformer processes text tokens and raw image patches, with no pretrained vision tower, text encoder, or decoder. (1/N 🧵) https://t.co/0SFjprphH8 21.6K views · 168 likes · 30 reposts · 11 replies Open on X →

Rispetto ad account della stessa dimensione

16 post degli ultimi 90 giorni, accanto alla fascia di under 10K follower. raggiunge meno persone dei pari, ma le coinvolge molto di più.

Visualizzazioni mediane382questo account4 175mediana per under 10K

Copertura, %24.44%questo account250.82%mediana per under 10K

Interazione, %3.80%questo account1.41%mediana per under 10K

Metrica	Questo account	Mediana per under 10K	Rapporto
Visualizzazioni mediane per post	382	4 175	0.09×
Copertura (visualizzazioni ÷ follower)	24.44%	2.5× audience	0.10×
Tasso di interazione	3.80%	1.41%	2.70×

Altri account di questa fascia → Confronta con un altro account → Come sono costruiti questi parametri →

Growth & engagement

How the posts we collected actually performed: views and reaction rate post by post, what the audience did with them, and where the follower count goes.

Views per post

7763 Sep

485

433

389

1.3K

336

1.2K

286

271

274

244

376

232

242

Last 14 collected posts, oldest on the left. The scale is logarithmic: one post can outrun the rest a hundred times over.

Engagement rate per post

2.32%3 Sep

3.30%

3.93%

3.34%

1.42%

4.76%

1.69%

5.59%

5.17%

6.57%

5.33%

3.72%

3.88%

4.13%

Reactions — likes, reposts, replies and quotes — divided by views. Median for under 10K accounts is 1.41%.

What the audience does

Likes68.5%382 in total

Reposts5.7%32 in total

Replies4.7%26 in total

Bookmarks21.1%118 in total

Share of every reaction we collected for this account. Replies mean argument, reposts mean endorsement, bookmarks mean the post was worth keeping.

The follower curve appears once this account has two daily snapshots — we take one a day, and this one is on its first.

Ultimi post

Rispetto ad account della stessa dimensione

Growth & engagement

Views per post

Engagement rate per post

What the audience does

Account simili