---
license: other
language:
- en
library_name: transformers
tags:
- RLHF
- Nexusflow
- Athene
- Chat Model
---
# Athene-V2-Chat-72B: Rivaling GPT-4o across Benchmarks
Nexusflow HF - Nexusflow Discord
We introduce Athene-V2-Chat-72B, an open-weights LLM that rivals GPT-4o across benchmarks. It is trained through RLHF based off Qwen-2.5-72B.
Athene-V2-Chat-72B excels in chat, math and coding. Its sister model, [Athene-V2-Agent-72B](https://huggingface.co/Nexusflow/Athene-V2-Chat), surpasses GPT-4o in complex function calling and agent applications.
Benchmark performance:
- **Developed by:** The Nexusflow Team
- **Model type:** Chat Model
- **Finetuned from model:** [Qwen 2.5 72B](https://huggingface.co/Qwen/Qwen2.5-72B-Instruct)
- **License**: [Nexusflow Research License](https://huggingface.co/Nexusflow/Athene-V2-Chat/blob/main/Nexusflow_Research_License.pdf)
- **Blog**: https://nexusflow.ai/blogs/athene-V2
## Usage
Athene-V2-Chat uses the same chat template as Qwen 2.5 72B. Below is an example simple usage using the Transformers library.
```Python
import transformers
import torch
model_id = "Nexusflow/Athene-V2-Chat"
pipeline = transformers.pipeline(
"text-generation",
model=model_id,
model_kwargs={"torch_dtype": torch.bfloat16},
device_map="auto",
)
messages = [
{"role": "system", "content": "You are an Athene Noctura, you can only speak with owl sounds. Whoooo whooo."},
{"role": "user", "content": "Whooo are you?"},
]
terminators = [
pipeline.tokenizer.eos_token_id,
pipeline.tokenizer.convert_tokens_to_ids("<|end_of_text|>")
]
outputs = pipeline(
messages,
max_new_tokens=256,
eos_token_id=terminators,
do_sample=True,
temperature=0.6,
top_p=0.9,
)
print(outputs[0]["generated_text"][-1])
```
We found that by adding system prompts that enforce the model to think step by step, the model can do even better in math and problems like counting `r`s in strawberry. For fairness consideration we **do not** include such system prompt during chat evaluation.
## Acknowledgment
We would like to thank the [LMSYS Organization](https://lmsys.org/) for their support of testing the model. We would like to thank Meta AI and the open source community for their efforts in providing the datasets and base models.