Said Taghadouini

pixels & tokens @ lighton / vit.cpp, modernbert, lightonocr

[Said Taghadouini @staghado]

🚀 LightOnOCR-2-1B 🦉 is out, a major update to LightOnOCR.
1B parameters, end-to-end multilingual OCR, and it beats models 9× larger on OlmOCR-Bench while being much faster.
PDF/page in, clean ordered Markdown out, with optional image localization (bbox variants).

[Said Taghadouini @staghado]

@AmelieTabatta @LightOnIO 🚀

[Amélie Chatelain @AmelieTabatta]

Do you like the open-source models we keep shipping at @LightOnIO? 👀
Now you can actually *build* with them!!

We're launching LightOn Console 🎮: three endpoints (Parse, Extract, Search) so you can run our models on your own documents without building the plumbing yourself!
🧵

[LightOn @LightOnIO]

Today, we're introducing LightOn Console.

⚙️ Three endpoints:
/Parse any documents
/Extract structured data
/Search enterprise knowledge with citations

🔌 Built-in connectors. MCP-ready. Governance enforced at the chunk level.

No infrastructure. No pipeline maintenance. No dedicated retrieval team required.
Make your enterprise knowledge agent-readable now!

Read the launch announcement: https://t.co/LcxXqyOgo5

Test it now: https://t.co/RNJQKEHzQ2

[Said Taghadouini @staghado]

@gabriberton same thing was in fuyu iirc, with enough compute no need for a vit but vlms starting from one get a head start.

[Said Taghadouini @staghado]

@vanstriendaniel qwen3.5 series is strong and fast!

[Amélie Chatelain @AmelieTabatta]

I still love seeing my old lecture notes as demo images:D

[Said Taghadouini @staghado]

@antoine_chaffin 👨‍🍳

[Said Taghadouini @staghado]

@adithya_s_k fyi the arabic samples are just garbled chars, something went wrong during rendering 😀

[Said Taghadouini @staghado]

character error rate is not a valid metric for full page ocr, thanks!

[Said Taghadouini @staghado]

@skalskip92 if qwen3.5 122B-A10B is only 5 tests away from gemini 3,5 then the 397B-A17B should be even closer.
plus it completely destroys everyone on speed?

[Said Taghadouini @staghado]

@RfK_001 FlexiViT did train with different patch sizes w/ weight sharing. But it was fixed resolution ViTs, i am not sure it's the same if you vary input resolution as they become very similar(either make the image bigger or make the patch smaller).

[Said Taghadouini @staghado]

@hu_yifei ig with thinking(default) it would figure it out easily

[Said Taghadouini @staghado]

ppl using this for parsing are cooked!

[INIYSA @lafaiel]

Gemini 3.5 is 30x more expensive than 1.5

[Said Taghadouini @staghado]

@hu_yifei is this one image or two separate images?

[Said Taghadouini @staghado]

@hu_yifei is it significantly better than 3/3.1 flash for parsing? tried the 3 today on some adversarial synth data and it was very decent.

[Said Taghadouini @staghado]

it still "fixes" some typos, a VLM is a VLM after all!

[Said Taghadouini @staghado]

gemini flash 3, i was not familiar with your game!
ps: sample size = 1 for now but good start

[Said Taghadouini @staghado]

@antoine_chaffin hahah probably not, nah i think it's fine for prompting purposes!

[Said Taghadouini @staghado]

created a simple tool to help me in my daily workflow: hit a hotkey, talk, the transcript pastes itself into Claude Code/Codex/wherever I'm typing. fully local, runs on my mac.

[Said Taghadouini @staghado]

try it
https://t.co/7yFoPfYHIL

[Said Taghadouini @staghado]

running @IBM Granite Speech 4 via llama.cpp @ggml_org.
credits to @RfK_001 for adding upstream support for this model.