China DeepSeek (DeepSeek) is updated to the next generation of AI history, an experimental version "DeepSeek-V3.2-Exp".
DeepSeek introduces the latest DeepSeek-V3.2-Exp through the Hugging Face platform, which uses a new technology called DeepSeek Sparse Attention (DSA) as a mid-step step towards a new generation of architectures.
DeepSeek shocked Silicon Valley this year with the inference model DeepSeek-R1, and planned to launch new products to strengthen its leading position in China's AI industry. DeepSeek-V3.2-Exp adds a sparse attention mechanism, namely DSA, to the DeepSeek-V3.1-Terminus basis, to improve and verify the training and reasoning efficiency of long texts.
DeepSeek-V3.2-Exp uses DSA to highlight DeepSeek to find ways to reduce training costs and difficulties, especially when it is restricted from obtaining NVIDIA chip resources. DeepSeek founder Liang Wen-hyun co-wrote this year on this topic with others, describing how developers combine software innovation with post-prepared hardware to reduce computing needs.
Reducing costs makes DeepSeek's model competitive, and DeepSeek cuts API prices in half or even more, joining other Chinese new founders to attract developers with significant price cuts. It is worth mentioning that Chinese AI chip manufacturer Hua announced to Cambrian that its products will support the latest model updates of DeepSeek.
DeepSeek also stated that the latest model supports FP8 accuracy and is working hard to support BF16. Models deal with millions of values, using smaller formats like FP8 and BF16 to balance speed and precision, making it easier to run large language models on constrained hardware. Although the FP8 is not very standard, it is useful for many AI tasks, and the BF16 is considered more accurate when training models.
💻 API Update
🎉 Lower costs, same access! 💰 DeepSeek API prices drop 50%+, effective immediately.
🔹 For comparison testing, V3.1-Terminus remains available via a temporary API until Oct 15th, 2025, 15:59 (UTC Time). Details: https://t.co/3RNKA89gHR 🔹 Feedback welcome:… pic.twitter.com/qEdzcQG5bu
— DeepSeek (@deepseek_ai) September 29, 2025
â–² Use DSA to improve long text performance and reduce computing costs.
Extended reading: DeepSeek Ambitions are still there, launching new models to continue fighting OpenAI Far below the competition opponent! China DeepSeek announces R1 model training cost only US$294,000 China DeepSeek launches V3.1 model to prepare domestic chips The main reason for the delay of the new R2 model! DeepSeek tried to fail because the chip was still up to NVIDIA DeepSeek-R1 The science behind the model, AI reasoning technology has been published in the journal Nature