DeepSeek open-sourced DeepSeek-R1 under the MIT license alongside dense distilled 1.5B–70B models, demonstrating open RLVR/GRPO reasoning competitive with proprietary frontier models and driving immediate optimization across vLLM and SGLang.
DeepSeek
Also known as: DeepSeek AI · DeepSeek-V3 · DeepSeek-R1
Plain English
A Chinese AI research lab that shook the global AI industry by releasing MIT-licensed open-weight models matching top proprietary labs at a fraction of reported training and inference costs, while publishing full architectural blueprints.
Technical Overview
DeepSeek pioneered Multi-head Latent Attention (MLA) for compact KV Cache, fine-grained Mixture of Experts (MoE) with auxiliary-loss-free load balancing, native FP8 mixed-precision training, Multi-Token Prediction (Speculative Decoding), and verifiable-reward RL (GRPO) in Post-Training (SFT, RLHF, DPO & GRPO)—releasing open Safetensors weights on Hugging Face.
Chronological Entries (2)
DeepSeek released DeepSeek-V3 (671B total / 37B active Mixture of Experts (MoE)), pairing MLA KV Cache compression, FP8 training, and Multi-Token Prediction—setting a new cost-efficiency baseline for frontier open-weight LLMs.
Direct Relations (4)
- ImplementsMixture of Experts (MoE)
- ImplementsPost-Training (SFT, RLHF, DPO & GRPO)
- ImplementsKV Cache
- UsesHugging Face