@abacaj Any success with granite docling? Was having AWFUL results but apparently needs really high DPI to get something usable.

@rsdgpt Poor results. Try olmocr https://huggingface.co/allenai... or dotsocr

@abacaj Thoughts on moondream?

@JsonBasedman I don’t use it but I like what they are doing. If you need document specific you need to fine tune a model for layout / ocr

@abacaj i use ColQwen + new Qwen3-4B-VL . it is pretty good, i wonder what kind of documents you are using? Unless you are feeding A LOT of pages, i think what i have with Qwen3VL is perfect

@abacaj Haven’t OSSed our models (perhaps in the future we might) but our models are getting pretty good (VLM + layout) https://x.com/ishaank1999/stat...

@abacaj ColPali?

@abacaj Not sure how a big lab hasn’t come in and crushed this with a dedicated cheap OCR model/API. Mistral tried but not enough reach? Or maybe it’s just too niche to bother for the big labs @OfficialLoganK ? Any chance of something like this?

@abacaj The problem is that often it is very vision heavy and not as language heavy. As you scale vlm's the vit, siglip, etc. Vision backbone stays the same size and just the language portion grows. Total waste to use a 70b llm and tie it to the same 300m vision encoder that the 2b uses

@abacaj These were hugely bottlenecked early on by resolution. When llava and gpt4o and co were just hitting the scene their puny resolution made it so that they were functionally blind to documents. It has gotten so much better in only a couple years. Still lots of progress to be made.

@abacaj https://x.com/ycombinator/stat... What about Nanonets-OCR2-3B?

18 months ago, @karpathy set a challenge: "Can you take my 2h13m tokenizer video and translate [into] a book chapter". We've done it! It includes prose, code & key images. It's a great way to learn this key piece of how LLMs work. https://www.fast.ai/posts/2025...

an RLM is just an LLM that has an RLM as a tool

Announcing the completely reimagined vLLM TPU! In collaboration with @Google, we've launched a new high-performance TPU backend unifying @PyTorch and JAX under a single lowering path for amazing performance and flexibility. 🚀 What's New? - JAX + Pytorch: Run PyTorch models on

MacStudio you ask? Apple Engineering's **actual** time spent on PyTorch support has't given me confidence that PyTorch Mac experience would get anywhere close to NVIDIA's any time soon, if ever. The Meta engineers continue to do a huge amount of heavy-lifting for improving the