DeepSeek releases DeepInfer inference framework as open source
DeepSeek has released the framework behind its production AI services, with support for DeepSeek-V4.1-Flash on Nvidia and Huawei hardware.
Company
DeepSeek is an artificial intelligence company developing AI models, developer tools, and infrastructure software. The company faces both rapid commercial expansion and heightened regulatory scrutiny across global markets. Currently, DeepSeek is advancing its technology stack through open-source releases and API expansions. It has open-sourced DeepInfer, the production inference framework behind its services, supporting its DeepSeek-V4.1-Flash model on Nvidia SM100 Blackwell GPUs and Huawei Ascend 950 chips. A single cluster serves 155 trillion tokens daily using this infrastructure, though cache management code remains unreleased. DeepSeek has also released TileLang, an open-source programming language designed for Huawei Ascend chips to offer an alternative to Nvidia CUDA. Developer offerings now include Search and Fetch APIs priced at $0.75 per 1,000 calls, alongside an updated V4.1-Flash model featuring improved coding capabilities. Financially, DeepSeek is nearing a funding round between $12 billion and $15 billion, driven primarily by CATL and Tencent, and plans to restructure for an initial public offering in early 2027. Simultaneously, DeepSeek encounters severe legal and regulatory obstacles. The United States Department of Homeland Security has ordered civilian government agencies to remove DeepSeek software and stop purchasing services built on its models over data security concerns. In China, officials are investigating the company over alleged data leaks to Anthropic. Furthermore, OpenAI previously accused DeepSeek of distilling its models, and the firm faces scrutiny over training on banned Nvidia Blackwell hardware.
8 stories · Updated
DeepSeek has released the framework behind its production AI services, with support for DeepSeek-V4.1-Flash on Nvidia and Huawei hardware.
The Department of Homeland Security ordered civilian agencies to remove DeepSeek and stop buying services built on its models.
Deepseek says its updated Flash model improves coding and agent tasks, while its Search and Fetch APIs are now available to developers.
Chinese AI startup Deepseek is nearing a $15 billion funding round and preparing for an initial public offering in early 2027.
Deepseek and Huawei have teamed up to release open-source programming tools for Huawei Ascend AI chips.
The Cyberspace Administration of China is investigating DeepSeek and Moonshot AI following Anthropic's claims that the labs routed sensitive data to U.S. servers.
A senior Trump administration official confirms DeepSeek used Nvidia Blackwell GPUs to train its next model, likely violating U.S. export controls.
OpenAI has formally asked U.S. lawmakers to act after alleging that DeepSeek systematically extracted ChatGPT outputs to train its own models.