Models · Updated 26 Oct, 11:51 pm IST
Transformers v5.21.0 adds MiniCPM-OCR, Flint and UltraVBert models
Why it matters for readers: You can read the linked docs and paper to understand how these models handle images and large language tasks.
- MiniCPM-OCR is a document-understanding model in OpenBMB's MiniCPM-V family pairing a SigLIP vision encoder with a Qwen3.5 language model, aimed at text-dense images like scans and receipts.1
- Flint is Recursive's mixture-of-experts (MoE) language model using hybrid attention and routing tokens to 6 of 256 experts plus an ungated shared expert.1
- UltraVBert extends ModernVBert by combining an UltraBERT text encoder with a SigLIP vision encoder, splitting pages into 512-pixel tiles and merging vision patches into tokens for retrieval.1
- The release notes include links to documentation and a paper for MiniCPM-OCR and provide technical details about architectures and attention mechanisms.1
Get a brief like this every morning
Uzha reads hundreds of sources and gives you the stories that matter for your work, with every source linked. Free.
Get started