DeepSeek’s V3.1 AI Model Targets Domestic Chip Use

DeepSeek’s latest AI model supports Chinese chip formats, adds hybrid inference modes, and updates its API pricing ahead of broader rollout.

Domestic Chip Optimization Signals Strategic Shift

Chinese AI startup DeepSeek has released an updated version of its flagship V3 model, known as DeepSeek-V3.1. The new iteration introduces support for UE8M0 FP8 precision, a format optimized for upcoming Chinese-made chips. This move aligns with broader efforts in China to reduce reliance on U.S. semiconductor technology amid ongoing export restrictions. While the company did not specify which domestic chips will be supported, the update suggests a strategic pivot toward compatibility with China’s emerging hardware ecosystem.

FP8, or 8-bit floating point, is a data format that enables faster and more memory-efficient AI processing. By adopting this format, DeepSeek aims to improve performance while preparing its models for integration with next-generation local chipsets. The update follows two earlier enhancements this year: a revision to the R1 model in May and a previous V3 upgrade in March. These developments reflect a steady cadence of technical refinement across DeepSeek’s core offerings.

Hybrid Inference and User Control Features

In addition to hardware compatibility, DeepSeek-V3.1 introduces a hybrid inference structure. This architecture allows the model to operate in both reasoning and non-reasoning modes, depending on the task at hand. Users can switch between these modes using a “deep thinking” toggle available on DeepSeek’s app and web platform. Both interfaces now run the updated version, offering more flexibility in how the model processes information.

The dual-mode setup may help balance performance and resource use, depending on whether a task requires complex reasoning or straightforward execution. Such adaptability could appeal to developers working across varied application domains. DeepSeek’s approach also reflects a growing trend in AI design: giving users more control over computational depth. This feature may become increasingly relevant as AI models are deployed in constrained environments.

API Pricing Update and Market Positioning

Starting September 6, DeepSeek will revise its API pricing structure. The API enables developers to integrate DeepSeek’s models into third-party applications and services. Adjustments to cost may reflect changes in infrastructure, usage patterns, or strategic positioning as the company expands its footprint. DeepSeek has previously drawn attention for offering competitive performance at lower operational costs compared to Western counterparts.

Earlier this year, the company made headlines by releasing models that rival systems like OpenAI’s ChatGPT. These models have been noted for their efficiency and affordability, particularly in enterprise and developer contexts. With V3.1, DeepSeek continues to refine its offerings while aligning with domestic technology goals. The update may also serve as a testbed for future deployments on Chinese hardware platforms.

FP8 Adoption Gains Momentum

The FP8 format used in DeepSeek-V3.1 is part of a broader industry shift toward lower-precision computing in AI. Major players like NVIDIA and Intel have also explored FP8 to boost performance and reduce energy consumption. While DeepSeek’s implementation is tailored for domestic chips, the underlying trend reflects a global push for more efficient AI architectures. If widely adopted, FP8 could reshape how models are trained and deployed across diverse hardware environments. As China accelerates its semiconductor development, AI firms like DeepSeek may play a key role in validating and scaling new chip technologies. The V3.1 model’s compatibility with future domestic processors positions it as a potential bridge between software innovation and hardware independence.