Gate News message, April 24 — DeepSeek V4-Pro and DeepSeek V4-Flash were officially released and open-sourced on April 24, with context processing length significantly expanded from 128K to 1M, representing nearly a 10-fold capacity increase. Huawei Computing announced that its Ascend supernode products fully support DeepSeek V4 series models through close collaboration between chip and model technologies.
Huawei Ascend 950 achieves high-throughput, low-latency DeepSeek V4 model inference deployment through fused kernel and multi-stream parallelism techniques to reduce Attention computation and memory access overhead. For DeepSeek V4-Pro with 8K input, Ascend 950 achieves approximately 20ms TPOT with 4,700 TPS single-card Decode throughput; for DeepSeek V4-Flash under 8K input, it reaches approximately 10ms TPOT with 1,600 TPS throughput. Ascend A3 supernode series also achieves full compatibility, with training reference implementations provided for rapid fine-tuning. Based on Ascend A3 64-card supernode with large EP mode, DeepSeek V4-Flash achieves over 2,000 TPS single-card Decode throughput in 8K/1K input-output scenarios using vLLM inference engine. Huawei’s full Ascend A2, A3, and 950 product lines support both DeepSeek V4-Flash and V4-Pro.
Huawei Cloud announced first-mover compatibility with DeepSeek V4, providing developers with one-click API token services through its MaaS platform. Huawei Cloud optimized system layer, operator layer, and cluster layer capabilities to ensure rapid model adaptation and high-performance deployment. Enterprises including Kingsoft WPS and 360 have already integrated DeepSeek’s new model via Huawei Cloud.
Cambricon also announced Day 0 compatibility with DeepSeek V4-Flash and V4-Pro based on the vLLM inference framework, with adaptation code open-sourced to the GitHub community. Cambricon previously achieved first-mover adaptation when DeepSeek V3.2 was released last year, having conducted deep software-hardware collaborative performance optimization on DeepSeek series models.
Disclaimer: The information on this page may come from third parties and does not represent the views or opinions of Gate. The content displayed on this page is for reference only and does not constitute any financial, investment, or legal advice. Gate does not guarantee the accuracy or completeness of the information and shall not be liable for any losses arising from the use of this information. Virtual asset investments carry high risks and are subject to significant price volatility. You may lose all of your invested principal. Please fully understand the relevant risks and make prudent decisions based on your own financial situation and risk tolerance. For details, please refer to
Disclaimer.
Related Articles
GPT-5.5 Returns to Cutting Edge in Coding, But OpenAI Switches Benchmarks After Losing to Opus 4.7
Gate News message, April 27 — SemiAnalysis, a semiconductor and AI analysis firm, released a comparative benchmark of coding assistants including GPT-5.5, Claude Opus 4.7, and DeepSeek V4. The key finding: GPT-5.5 marks OpenAI's first return to the cutting edge in coding models in six months, with S
GateNews2m ago
Google DeepMind Executive: Every AI Product Company Should Build Custom Benchmarks
Gate News message, April 27 — Logan Kilpatrick, senior product manager at Google DeepMind and product lead for Google AI Studio, stated on X that every company building AI-based products should establish its own custom benchmarks to measure AI model performance. He described this as a method to
GateNews1h ago
MediaTek secures a major order for Google’s 8th-generation TPU! “ASIC” market sentiment catches on, benefiting three concept stocks
MediaTek has entered Google’s 8th-generation TPU training chip supply chain, responsible for the I/O Die design, and adopts TSMC’s N3P and CoWoS-S packaging, showing that it has risen to a high-end AI ASIC design level. It is expected that this year ASIC revenue will surpass $1 billion, and TPU shipment volume could reach tens of millions of units by 2027. Taiwan’s supply chain companies such as Jingyuan Electronics, Chip Implementation Testing, and Hon Hai will also benefit, as the market gradually shifts toward future division of labor and collaboration, as well as chiplet-based heterogeneous integration.
ChainNewsAbmedia1h ago
Singtel's RE:AI Partners with Mistral AI to Develop AI Infrastructure in Singapore
Gate News message, April 27 — Singtel Group announced that RE:AI, its Singapore-based digital infrastructure unit, has partnered with Mistral AI to support AI capabilities for industries in Singapore.
The partnership will focus on co-developing AI infrastructure and solutions for financial
GateNews1h ago
He Yi's YZi Labs Invests in Chinese AI Large Language Model Company
Gate News message, April 27 — He Yi, alongside Changpeng Zhao's family office YZi Labs, has invested in a Chinese artificial intelligence large language model company, she disclosed during a private KOL gathering at the Hong Kong Web3 Carnival. The specific name of the invested company was not discl
GateNews1h ago
OpenAI Partners with MediaTek and Qualcomm to Develop AI Smartphone Processor, Foxconn Connector to Begin Production in 2028
Gate News message, April 27 — OpenAI is collaborating with MediaTek and Qualcomm to develop smartphone processors, according to supply chain analysis by Tianfeng International Securities analyst Ming-Chi Kuo. Foxconn Connector (Luxshare Precision) will serve as the exclusive system design and
GateNews1h ago
