BEIJING, April 24 - Chinese startup DeepSeek on Friday published a preview release of V4, its next-generation artificial intelligence model family that has been adapted to run on Huawei’s most advanced Ascend AI chips. The release represents another step in efforts to build a more self-reliant domestic AI technology stack in China.
Model design and capabilities
DeepSeek describes V4 as engineered for use with agent-style frameworks, specifically naming Claude Code and OpenClaw as compatible systems. That alignment reflects an industry shift away from single-turn, prompt-driven chatbots toward agents capable of carrying out multi-step, complex tasks with less hands-on prompting.
V4 is offered in two configurations. The Pro variant is framed as the higher-end offering - more computationally intensive and costlier - while the Flash variant is presented as a lighter, lower-cost alternative. According to a technical paper released by DeepSeek alongside the preview, V4-Pro excels in agentic coding, world knowledge, STEM subject areas and competitive programming tests. In maximum reasoning mode, DeepSeek asserts that Pro outperforms all existing open-source models, even as it concedes that it lags some frontier closed-source systems such as Google’s Gemini 3.1 Pro and OpenAI’s GPT-5.4 in certain dimensions.
"DeepSeek-V4-Pro Max ... redefines the state of the art for open models, outperforming its predecessors in core tasks," the company said in its accompanying paper.
Flash is characterized as delivering similar reasoning performance in some contexts while operating faster and at lower cost than Pro. However, DeepSeek notes that Flash offers weaker world knowledge and reduced effectiveness on demanding agent-based tasks compared with Pro. Both Pro and Flash support a 1-million-token context window - an expansion matching the larger context capacity the company introduced with V3 in February. DeepSeek also highlights that V4’s architecture includes design elements intended to reduce compute and memory overheads when using very long contexts.
Adaptation for Huawei hardware
A notable difference from earlier DeepSeek releases is that V4 has been adapted for Huawei’s Ascend AI chips. DeepSeek did not make its new model available to U.S. chipmakers for performance tuning; instead, it provided early access to domestic firms including Huawei, even though the company had previously worked closely with Nvidia’s technical staff. Within hours of the V4 preview, Huawei said its Ascend 950-based supernode clusters fully support the DeepSeek-V4 series and that Ascend chips were used in part of V4-Flash’s training.
"Through close technical collaboration ... the entire Ascend supernode product line now supports the DeepSeek-V4 series models," Huawei said.
DeepSeek’s earlier V3 and R1 models were trained on Nvidia hardware. The company did not state whether Nvidia chips were used in training any components of V4.
Implications for domestic self-sufficiency and remaining limits
Analysts see the Huawei adaptation as a practical step toward reducing dependence on foreign AI infrastructure. Lian Jye Su, chief analyst at tech research firm Omdia, said the effort demonstrates DeepSeek models can deliver comparable performance across both Huawei and Nvidia hardware ecosystems. He added that the popularity of DeepSeek domestically encouraged Huawei to optimize the model for its silicon, lowering barriers for Chinese developers to build AI applications using fully domestic toolchains.
He cautioned, however, that Huawei still trails Nvidia on some technological fronts and that shifting developers away from Nvidia’s broader ecosystem remains challenging. "DeepSeek’s pivot reveals real, tangible progress toward AI infrastructure self-sufficiency," he said.
DeepSeek also faces practical compute constraints resulting from U.S. export controls on Nvidia chips and related chipmaking equipment. The company said the Pro tier can cost as much as 12 times more than Flash because of "constraints in high-end compute capacity," which in turn has limited the current availability of Pro services. DeepSeek said Pro pricing could fall significantly once Huawei Ascend 950 supernodes are deployed at scale in the second half of the year.
Summary
DeepSeek’s V4 preview brings an open-source model family optimized for agentic tasks to Huawei Ascend hardware, offering Pro and Flash variants with a 1-million-token context window. While DeepSeek touts Pro’s gains over prior open models, it acknowledges gaps versus the leading closed-source systems. Huawei confirms Ascend 950 support and partial use in Flash training, and analysts view the move as tangible progress toward a more domestic AI infrastructure - though hardware and export constraints continue to limit high-end compute availability.
Key points
- V4 supports agent frameworks such as Claude Code and OpenClaw and is offered in Pro and Flash variants.
- Both variants support a 1-million-token context window; V4 architecture aims to lower compute and memory costs for long-context use.
- Huawei confirmed Ascend 950 supernode support and said its chips were used in part of V4-Flash’s training, underscoring efforts toward domestic AI infrastructure.
Risks and uncertainties
- High-end compute constraints tied to export controls could keep Pro capacity scarce and pricing elevated - impacting cloud, AI services and enterprise adopters that need large-scale inference and training capacity.
- Huawei remains technologically behind Nvidia in some areas, which may limit the speed at which developers and enterprises migrate away from Nvidia-centered ecosystems.
- DeepSeek has not clarified the full mix of hardware used to train V4, leaving questions about cross-platform performance tuning and benchmarking.