DeepSeek LLM 67B

DeepSeek LLM 67B is the large base model released by DeepSeek on November 29, 2023, trained from scratch on 2 trillion tokens of English and Chinese text. It shipped in Base and Chat variants, with DeepSeek reporting a 73.78 HumanEval score and claiming it beat Meta's Llama 2 70B on several evaluations.

License nuance: the weights are distributed under DeepSeek's custom Model License, not MIT — the code in the repo is MIT, but the model license is a separate commercial-friendly agreement that is not OSI-open. Like most base models of its era, it lacks instruction tuning and safety alignment out of the box, so production use needs further fine-tuning and guardrails.