DeepSeek — Latest News and Analysis

DeepSeek has entered talks to raise $300M or more at a $10B valuation, its first outside round amid a delayed V4 and talent drain.

OpenAI, Anthropic, and Google have joined forces through the Frontier Model Forum to detect and counter unauthorized AI model copying by Chinese firms.

Microsoft has launched an Africa AI offensive, targeting 3 million users for training in 2026 while partnering with MTN to counter DeepSeek's growing dominance.

NVIDIA has detailed how its Dynamo inference framework and Brev platform tackle planetary-scale AI inference, cutting per-token costs 35x on GB200 hardware.

OpenAI has warned U.S. lawmakers that DeepSeek has been training its AI models through unauthorized distillation methods and circumvention tactics in escalating IP dispute.

Zhipu AI has become the first Chinese company to train a major AI model entirely on Huawei's domestic chips, releasing the open-source GLM-Image system.

DeepSeek has unveiled plans for a multimodal AI search engine processing text, images, and audio, challenging Google's keyword-based dominance with agents.

DeepSeek has released its OCR 2 model with semantic reasoning architecture that abandons traditional scanning, achieving 91.09% on benchmarks.

Google researchers have discovered that AI reasoning models like DeepSeek-R1 and QwQ-32B simulate internal debates between conflicting viewpoints.

How DeepSeek abandoned attempts to train its R2 AI model on Huawei chips after persistent failures, switching back to Nvidia hardware despite Beijing's push for domestic semiconductors.

Microsoft has reported that Chinese AI platform DeepSeek has captured 89% of China's AI market and is gaining market share in Africa and sanctioned countries at 2-4 times the rate of other regions.

DeepSeek has expanded its R1 whitepaper by 60 pages to disclose training secrets, clearing the path for a rumored V4 coding model launch.

MiniMax has raised $619 million in its Hong Kong IPO at a $6.5 billion valuation, joining Zhipu AI as Chinese firms beat OpenAI to public markets.

DeepSeek has introduced Manifold-Constrained Hyper-Connections (mHC), a novel architecture that stabilizes AI training and enables scaling on constrained hardware.

Nvidia has developed a software "digital leash" to track AI GPU locations via latency, responding to a $160M smuggling bust and DeepSeek's Blackwell leak.

DeepSeek has reportedly acquired 2,000 banned Nvidia Blackwell chips, overshadowing a $160 million DOJ bust of smugglers moving older H100 hardware.

Moore Threads has surged over 400% on its Shanghai debut, raising $1.1 billion to challenge Nvidia despite US sanctions and profitability concerns.

DeepSeek has released V3.2-Speciale, a high-compute reasoning model claiming Gold Medal math performance, but restricts access to a temporary 14-day API window.

DeepSeek has released DeepSeekMath-V2, an open-weight AI that has matched OpenAI and Google's Gold Medal performance at the IMO 2025.

With the "easy money" era of AI scaling coming to an end, Google’s 1,000x mandate for the "Age of Inference" signals a capital-intensive war that threatens the very foundations of OpenAI and Nvidia.

DeepSeek has released DeepSeek-OCR, an open-source AI model that uses 'optical compression' to efficiently process long documents after its R2 model was delayed.

OpenAI has banned state-backed accounts from China and North Korea for using its AI for surveillance, phishing, and cyber espionage, a new report reveals.

Google has declined to fix a critical 'ASCII smuggling' vulnerability in its Gemini AI, leaving Google Workspace users exposed to hidden attacks that can spoof identities and poison data.

Chinese AI lab DeepSeek has released DeepSeek-V3.2-Exp, an experimental open-source model testing a new, more efficient sparse attention mechanism.

In a major pivot, Tesla is integrating AI from China's Deepseek and Bytedance for its cars, sidelining Elon Musk's own Grok amid performance concerns.

Chinese AI firm DeepSeek's R2 model launch is delayed again after hitting technical hurdles training on Huawei's Ascend chips, a major setback for China's tech ambitions.

xAI's Grok 4 has dominated Day 1 of Google's Kaggle Game Arena, a new chess tournament testing the strategic reasoning of top LLMs from Google, OpenAI, and others.

Chinese AI startup Z.ai has released its open-source GLM-4.5 models, escalating the AI price war with rival DeepSeek and focusing on advanced agentic capabilities.

NVIDIA has released OpenReasoning-Nemotron, a family of open-source AI models that set new reasoning records by distilling them from China's DeepSeek R1-0528.

Former Intel CEO Pat Gelsinger launches the Flourishing AI (FAI) benchmark via Gloo, a new test grading AI on human values like character, meaning, and faith.