DeepSeek's DSpark module accelerates AI inference by using a lightweight draft model for candidate responses, verified in batches by a larger model, and employs semi-autoregressive generation and confidence-based scheduling for balanced speed and quality.
Chinese AI firms are changing pricing strategies due to competition from DeepSeek's V4 models, with Xiaomi slashing API costs by 99% for its MiMo-V2.5 model
DeepSeek V4 yapay zeka modelinin piyasaya sürülmesiyle Huawei Ascend 950 çiplerine olan küresel ve yerel talep hızla yükseldi. ByteDance, Tencent ve Alibaba gibi Çin'in önde gelen teknoloji devleri, yeni nesil yapay zeka iş yüklerini yönetebilmek amacıyla Huawei ile stratejik görüşmelere başladı. Ascend 950 işlemci ailesi, DeepSeek V4-Flash ve V4-Pro modelleriyle tam uyumlu çalışacak şekilde optimize edildi.
The potential emergence of a new DeepSeek spokesman has sparked intrigue as speculation over the whereabouts of founder and CEO Liang Wenfeng continues, more than a year after his last public appearance. While the Hangzhou start-up’s latest V4 model did not make the same waves as its breakout moment a year earlier, the much-anticipated release still grabbed headlines through a collaboration with domestic tech giant Huawei Technologies and its remarkably low prices. For the traditionally...
ABD Dışişleri Bakanlığı, Çinli şirketlerin ABD yapay zekâ modellerini 'damıtma' yoluyla kopyaladığı iddialarıyla ilgili küresel bir uyarı kampanyası başlattı. Diplomatik yazı, dünyanın dört bir yanındaki temsilciliklere gönderildi. OpenAI da DeepSeek'in modellerini çaldığını iddia etmişti. DeepSeek V4 modelini Huawei çip teknolojisine uyarladı.
DeepSeek's V4 Pro model ranked second globally among open-source AI models, trailing Moonshot AI's Kimi K2.6, according to benchmark firm Artificial Analysis. While V4 Pro showed improvement over its predecessor V3.2, it still lagged behind top competitors, highlighting challenges facing Chinese AI firms racing to close the gap with the US amid intensifying competition and computing power constraints.
Huawei Technologies’ newest Ascend 950PR and 950DT chips enjoyed “day zero” adaptation to DeepSeek’s latest V4 model, the Shenzhen-based tech giant said in a livestream on Friday, just hours after the model’s release. The hardware-software collaboration between the two Chinese firms underscores the progress China has made in tech self-reliance, a top national priority for Beijing amid US efforts to block its access to advanced semiconductor chips and chipmaking equipment. During the livestream...
Chinese AI company DeepSeek has unveiled a preview of its V4 model, claiming performance parity with leading US systems. The release emphasizes improved coding capabilities and compatibility with domestic Huawei chips, following previous controversies regarding training costs and data usage.
DeepSeek, Cuma günü yeni V4 modelinin ön sürümünü tanıttı. Pro versiyonu, açık kaynaklı modeller arasında en yüksek performansı gösterirken yalnızca Google'ın Gemini-Pro-3.1'in gerisinde kalıyor. Şirket, 20 milyar doların üzerinde değerleme ile fon toplama hedefinde; Alibaba ve Tencent hisse almak için görüşmeler yürütüyor.
Chinese AI startup DeepSeek is hiring for data center roles in Inner Mongolia, signaling an expansion of its computing infrastructure ahead of the V4 model release. This move raises questions about its sourcing of AI chips.
