Grok-1
Grok-1 overview
#TL;DR（中文）
- 是 xAI 发布的 open-weightcode
Grok-1模型之一（base model checkpoint），更偏研究/对照用途，而不是开箱即用的 chat agent。codeMoE - 你可以把它当成一个“基础模型样本”：需要额外的 instruction tuning / alignment 才更适合对话与应用落地。
- 如果要用于工程，建议先做：任务级 、安全策略（guardrails）、以及必要的 fine-tuning 或 prompt constraints。code
evaluation
#中文导读（术语保留英文）
阅读这页建议关注：
- MoE 的 inference 特性（active weights 比例）
- pretraining cutoff（对知识时效的影响）
- benchmark 结果的定位（更适合对比，不代表你的任务一定表现一致）
#Original (English)
Grok-1 is a mixture-of-experts (MoE) large language model (LLM) with 314B parameters which includes the open release of the base model weights and network architecture.
Grok-1 is trained by xAI and consists of MoE model that activates 25% of the weights for a given token at inference time. The pretraining cutoff date for Grok-1 is October 2023.
As stated in the official announcement, Grok-1 is the raw base model checkpoint from the pre-training phase which means that it has not been fine-tuned for any specific application like conversational agents.
The model has been released under the Apache 2.0 license.
#Results and Capabilities
According to the initial announcement, Grok-1 demonstrated strong capabilities across reasoning and coding tasks. The last publicly available results show that Grok-1 achieves 63.2% on the HumanEval coding task and 73% on MMLU. It generally outperforms ChatGPT-3.5 and Inflection-1 but still falls behind improved models like GPT-4.
Grok-1 was also reported to score a C (59%) compared to a B (68%) from GPT-4 on the Hungarian national high school finals in mathematics.
Check out the model here: https://github.com/xai-org/grok-1
Due to the size of Grok-1 (314B parameters), xAI recommends a multi-GPU machine to test the model.
