Podcast Episode
The model achieved a ninety-one percent score on the OmniDocBench benchmark while using remarkably few computational resources, compressing complex document pages into just two hundred and fifty-six to one thousand one hundred and twenty visual tokens.
The V4 release would integrate new Engram conditional memory technology for efficient retrieval across extremely long contexts, potentially running on consumer-grade hardware including dual RTX 4090 graphics cards.
DeepSeek Expands AI Empire with Search Engine, OCR Model, and V4 Announcement
January 29, 2026
Audio archived. Episodes older than 60 days are removed to save server storage. Story details remain below.
Chinese AI startup DeepSeek is rapidly expanding beyond large language models, announcing plans to build a multimodal AI search engine to rival Google and OpenAI while simultaneously releasing its advanced OCR 2 model and preparing to launch V4 in mid-February.
DeepSeek Takes Aim at Google and OpenAI
Chinese artificial intelligence powerhouse DeepSeek is making its most ambitious move yet, hiring specialists to build a multimodal AI search engine capable of processing text, images, and audio queries across multiple languages. The Hangzhou-based company's job postings this month signal a direct challenge to both Google and OpenAI in the lucrative search market.New OCR Model Mimics Human Reading
Alongside its search ambitions, DeepSeek released its optical character recognition model, DeepSeek-OCR 2, on Tuesday. The three-billion-parameter model introduces a revolutionary visual causal flow architecture that mimics how humans naturally interpret complex document layouts, rather than scanning images in rigid patterns.The model achieved a ninety-one percent score on the OmniDocBench benchmark while using remarkably few computational resources, compressing complex document pages into just two hundred and fifty-six to one thousand one hundred and twenty visual tokens.
V4 Model Promises Coding Revolution
DeepSeek is also preparing to launch its next-generation V4 model around mid-February, with internal tests suggesting it may outperform competitors in coding capabilities. The model reportedly handles context windows exceeding one million tokens, enabling developers to process entire codebases in a single pass.The V4 release would integrate new Engram conditional memory technology for efficient retrieval across extremely long contexts, potentially running on consumer-grade hardware including dual RTX 4090 graphics cards.
Aggressive Talent Acquisition
The company has ramped up recruitment efforts, posting over forty technology positions on Chinese job platforms and ten new listings on LinkedIn. Roles span artificial general intelligence research, large language model development, and autonomous agent design, with DeepSeek emphasising its top-tier computing infrastructure as a competitive advantage in attracting global talent.Published January 29, 2026 at 10:43am
