UPDATE — NOVEMBER 2025: Since DeepSeek announced the V3.1 upgrade in August 2025—including support for upcoming Chinese-made chips, a new FP8 precision format, and hybrid inference modes—the company has issued additional improvements and clarifications about the model’s performance and ecosystem alignment.
On September 22, 2025, DeepSeek released V3.1-Terminus, a refined version designed to fix issues observed in early deployments. The Terminus update improves language consistency by reducing mixed Chinese-English outputs and minimizing rare-character errors—a key concern for enterprise and government users in China. It also boosts stability and accuracy for DeepSeek’s Code Agent and Search Agent across all access points: app, web client, mini program, and the developer API.
The V3.1 series now supports long-context processing up to 128K tokens, enabling substantially longer conversations, document analysis, and code handling. DeepSeek also expanded compatibility by aligning its API format with the Anthropic Claude API, making it easier for global developers to switch or integrate the model. In addition, the hybrid “thinking vs. non-thinking” inference mode introduced in August has been upgraded with faster reasoning throughput and more reliable tool-calling behavior. Pricing changes for the API—first scheduled for September 6—have since taken effect, though DeepSeek still has not publicly disclosed which domestic chip manufacturers its FP8 format is optimized for.
Taken together, these updates position V3.1 as one of the most aggressively iterated Chinese large-model platforms, tightly aligned with national goals to reduce reliance on U.S. chips while pushing model performance and efficiency beyond the initial August release.
ORIGINAL NEWS STORY:
DeepSeek Upgrades AI Model to Support China’s Domestic Chips Amid Tech Push
Chinese artificial intelligence startup DeepSeek has released an upgrade to its flagship V3 model, introducing compatibility with Chinese-made chips and faster processing speeds, Reuters reported.
The upgraded model, DeepSeek-V3.1, features a UE8M0 FP8 precision format designed to optimize performance for “soon-to-be-released next-generation domestic chips,” the company said in a WeChat post. The move underscores Beijing’s push to strengthen its semiconductor ecosystem and reduce reliance on U.S. technology amid Washington’s export restrictions.
DeepSeek has been making headlines this year as one of the few Chinese startups producing AI models that compete with Western systems like OpenAI’s ChatGPT, while touting lower operational costs. The latest update builds on recent improvements to its R1 model in May and an earlier V3 enhancement in March.
According to the company, the new version also introduces a hybrid inference structure, allowing the model to toggle between reasoning and non-reasoning modes. Users can switch modes using a “deep thinking” button on the firm’s app and web platform, which now run on V3.1.
In addition to the technical updates, DeepSeek announced that pricing for its application programming interface (API) — used by developers to integrate its AI capabilities into other apps and products — will be adjusted beginning September 6.
FP8, the 8-bit floating point format embedded in the V3.1 model, is designed to make AI systems more efficient by reducing memory usage while maintaining high-speed performance. However, DeepSeek did not disclose which domestic chipmakers or products the upgrade will support.
Need Help?
If you’re wondering how any AI systems, regulations and laws worldwide could impact you and your business, don’t hesitate to reach out to BABL AI. Their Audit Experts can address your concerns and questions while offering valuable insights.
