DeepSeek V4 has been released! 🐳

DeepSeek V4 has been released! 🐳 This is the largest open model, outperforming Kimi K2.6 in benchmarks and getting very close to closed models. It comes in two variants - Pro with 1.6T-A49B parameters and Flash with 284B-A13B. The models support a million tokens of context, and the new attention scheme allows reducing the KV cache by 10 times on a context of a million tokens. Price - $0.14/$0.28 per million tokens for Flash, like DeepSeek V2, and $1.74/$3.48 for Pro. The models are already available in chat, Instant is Flash, and Expert is Pro. A detailed technical report has been released with the model.