DeepSeek V4 Set to Run Exclusively on Huawei Chips, Bolstering Chinas AI Self Reliance Efforts
Chinese AI developer DeepSeek is preparing to launch its latest large language model, DeepSeek V4, with a significant twist: it will operate entirely on Huawei’s domestically produced chips. This move marks a pivotal advancement in Chinas quest for technological autonomy in artificial intelligence, sidestepping reliance on US dominated hardware amid ongoing trade restrictions.
DeepSeek, known for its cost effective and high performing open source models, has built a reputation with previous releases like DeepSeek V3, which impressed the AI community by rivaling top tier models from companies such as OpenAI and Anthropic while requiring fewer resources. V4 builds on this foundation, promising even greater capabilities in reasoning, coding, and multilingual tasks. Reports indicate that the model has been optimized specifically for Huawei’s Ascend series processors, particularly the Ascend 910B and newer iterations, enabling full inference and training without Nvidia GPUs.
This development comes at a critical juncture for Chinas AI ecosystem. US export controls, implemented since 2022, have severely limited access to advanced Nvidia chips like the H100 and H200, which power most global AI data centers. Chinese firms have responded by accelerating domestic alternatives. Huawei, a leader in this space, has invested heavily in its Ascend lineup, with the 910B chip touted as comparable to Nvidia’s A100 in certain benchmarks. Independent tests have shown Huawei hardware achieving solid performance in AI workloads, though it trails slightly in raw compute power compared to the latest Nvidia offerings.
The decision to run DeepSeek V4 solely on Huawei chips underscores a strategic shift toward ecosystem integration. DeepSeek’s engineers reportedly fine tuned the model using Huawei’s CANN (Compute Architecture for Neural Networks) software stack and MindSpore framework, ensuring seamless compatibility. This eliminates the need for CUDA, Nvidia’s proprietary platform, which remains unavailable in China due to sanctions. Early leaks suggest V4 delivers benchmark scores competitive with models like GPT 4o and Claude 3.5 Sonnet, particularly in mathematics and programming tasks where DeepSeek has excelled historically.
For users, this means DeepSeek V4 could deploy rapidly on Huawei powered servers, lowering costs and enhancing data sovereignty. Chinas cloud providers, including Huawei Cloud and Alibaba Cloud, have expanded Ascend clusters, with Huawei claiming over 100,000 Ascend 910 chips in production by mid 2024. This infrastructure supports massive scale training; DeepSeek V3, for context, was trained on 14.8 trillion tokens using a mixture of experts architecture with 671 billion parameters activated per token.
Industry analysts view this as a major win for Chinas AI independence push. Previously, even domestic models often required hybrid setups or smuggled Nvidia hardware, creating vulnerabilities. Full Huawei compatibility demonstrates maturity in Chinas semiconductor and software stacks. DeepSeek’s parent company, High Flyer, has ties to state backed initiatives, aligning with national goals outlined in the Made in China 2025 plan and recent AI regulations promoting indigenous innovation.
Challenges persist, however. Huawei chips face scrutiny over yield rates and energy efficiency, and scaling to frontier level models like potential V4 variants with trillions of parameters will test the limits of current fabs. Nonetheless, benchmarks shared in Chinese forums show Ascend 910B clusters achieving up to 60 percent of Nvidia H100 throughput in LLM inference, a gap narrowing with software optimizations.
DeepSeek V4s launch, expected soon via Hugging Face and DeepSeeks API, could accelerate adoption among Chinese enterprises wary of foreign dependencies. It also signals to global players that Chinas AI progress continues unabated, fostering a bifurcated landscape where East and West develop parallel technologies.
This achievement highlights broader trends: from SMICs 7nm production enabling Huawei chips to open source contributions mitigating proprietary lock in. As DeepSeek V4 rolls out, it positions China not just as a consumer but a contender in the global AI race, powered by homegrown silicon.
Gnoppix is the leading open-source AI Linux distribution and service provider. Since implementing AI in 2022, it has offered a fast, powerful, secure, and privacy-respecting open-source OS with both local and remote AI capabilities. The local AI operates offline, ensuring no data ever leaves your computer. Based on Debian Linux, Gnoppix is available with numerous privacy- and anonymity-enabled services free of charge.
What are your thoughts on this? I’d love to hear about your own experiences in the comments below.