On 20 October 2026, DeepSeek revealed a significant technological advancement with the publication of its V4.1-Flash inference engine. This announcement comes at a crucial time for the CareTech industry and the Biosphere, as it marks a turning point in how algorithms can be used to enhance well-being and health, as well as the preservation of the biosphere.

Version and Support

DeepInfer, the V4.1-Flash inference engine, is now available under Apache 2.0 licensing. It is designed to work on NVIDIA SM100 and Huawei Ascend 950 architectures. This version is limited to offline inference, meaning that pre-processing and decoding operations are not integrated.

Notable Features

  • Offline Inference : The engine is currently limited to offline inference, meaning it cannot handle external preprocessing and decoding.
  • Cache KV and Load Balancing: These features are internal and are not available in this version.
  • Pricing: DeepSeek has indicated a rate of $38 per GPU-hour, a significant increase compared to the theoretical revenue for March 2025, which was nearly triple. However, this rate remains somewhat incomparable and suggests an aggressive pricing strategy.

History of DeepSeek

It is interesting to note that DeepSeek had abandoned plans to launch its internal engine in April 2025 due to technological and commercial challenges. The recent publication of DeepInfer could be seen as a direct response to these challenges and an indication of the significant progress made by the company.

Technological Evolution

The release of DeepInfer as an offline inference engine represents a significant evolution compared to previous versions. The inclusion of offline inference functionality, even without integrated pre-processing and decoding, is a major step towards ease of use and adoption by new users.

CareTech

In the field of CareTech, offline inference could open new perspectives for health applications and medical data processing. Improving the performance of these inference engines can lead to faster and more precise solutions for diagnosis and personalized therapy.

Biosphere

For the Biosphere, offline inference could also play a crucial role. Modelling and predicting environmental phenomena, as well as analysing planetary health data, could benefit from an upgrade of the algorithms.

Challenges and Opportunities

Although offline inference is an advance, there are still many challenges to overcome. Total pre-processing/decoding separation is not yet implemented, which could limit certain applications. However, the significant increase in pricing suggests a strategy to valorise the technology, which could stimulate innovation and adoption.

Cet article est rédigé par Blaise Grenier, Journaliste reporter d'images et auteur. Découvrir le parcours de l'auteur et la charte éditoriale »