34 articles
DeepSeek's latest V4Lite checkpoint is showing improved benchmark performance, particularly in math and coding, as competition in the AI model landscape continues to intensify.
Grok's weekend traffic decline was the lowest among major generative AI tools, according to traffic data shared online. Competing platforms like ChatGPT and DeepSeek saw steeper drops.
DeepSeek has reportedly denied Nvidia and AMD access to its latest AI model, diverging from typical industry practice. The situation also includes unverified claims the model was trained on banned Nvidia chips.
DeepSeek's new "DualPath" paper challenges the standard prefill-centric approach to KV-cache loading, reporting up to 1.87x higher offline throughput and up to 1.96x more agent runs per second in online serving, driven by smarter scheduling and infrastructure design.
A 27-billion-parameter model is closing the gap on the Artificial Intelligence Index, challenging larger and more established AI systems across 10 benchmarks.
DeepSeek V3's Multi-Token Prediction (MTP) was applied solely during training to sharpen model representations, not to accelerate inference. The distinction highlights how training objectives and runtime optimizations serve very different purposes in large AI models.
Multiple flagship AI models are rolling out across leading platforms in the coming weeks. The packed release schedule signals how development timelines are tightening across the industry.
Nvidia's GB300 GPUs deliver breakthrough performance running DeepSeek models, with benchmarks showing up to 20x throughput improvements over previous-generation hardware in AI inference workloads.
Grok pulled ahead of DeepSeek for the first time ever in January, claiming the #3 spot among the world's most popular AI platforms.
DeepSeek AI has released DeepSeek-OCR 2, a new open-source visual-language model that improves optical character recognition and document understanding through a redesigned vision encoder for better reading order and OCR accuracy.
Grok just pulled ahead of DeepSeek in global website traffic for the first time, hitting 3.5% market share in mid-January 2026. The data from Similarweb shows how fast things can shift in the AI space.
Two Chinese AI labs are pushing language models forward through different but complementary breakthroughs—one reshaping how neural networks connect internally, the other revolutionizing how AI systems store and access information.
Microsoft data reveals stark geographic divides in DeepSeek usage, with adoption reaching 89% in China and over 40% in Russia while remaining minimal in Western markets where access restrictions appear to drive platform choices.
DeepSeek's new Engram architecture introduces a deterministic memory lookup system that slashes redundant computation in large AI models by letting them retrieve stored knowledge directly instead of recalculating everything from scratch.
DeepSeek just dropped a research paper on Engram—a new conditional and scalable memory module built to make large language models way more efficient. The timing's got everyone buzzing about what their next-gen model might look like.
Zhejiang High-Flyer Asset Management delivered average returns of 56.6% in 2025, highlighting a strong year for Chinese quant funds and strengthening funding capacity tied to DeepSeek.
A fresh report claims China's DeepSeek trained its upcoming AI model using Nvidia's Blackwell chips that were smuggled into the country, dodging U.S. export restrictions. The chips reportedly traveled through third countries before being taken apart and shipped into China piece by piece.
DeepSeek V3.2 now holds the highest score among open source AI models on the Cortex-AGI benchmark, with 38.2 %. It places sixth across all entries. The result shows that open source systems steadily gain strength in tasks that demand advanced reasoning.
DeepSeek just dropped two new AI models, with V3.2 Speciale beating GPT-5 High on key benchmarks while staying completely open-source. The release also shows strong performance against Gemini 3.0 Pro at way lower costs.
DeepSeek released an AI that reasons about mathematics. The program writes a proof, checks the proof and fixes any error without human help. It took the 2024 Putnam test plus earned 118 points from a possible 120. It also beat the top models in the main math contests.
We use cookies to improve your experience on our site and to show you relevant advertising. To find our more, read our privacy policy and cookie policy
Eseandre Mordi
Saad Ullah
Peter Smith
Usman Salis
Alex Dudov
Marina Lyubimova
Artem Voloskovets
