Welcome to the xGen-small family!
xGen-small (blog, arXiv) is an enterprise-ready compact LM that combines domain-focused data curation, scalable pre-training, length-extension, and RL fine-tuning to deliver long-context performance at predictable, low cost. This model release is for research purposes only.
Model Series
xGen-small comes in two sizes (4B and 9B) with two variants (pre-trained and post-trained):
| Model | # Total Params | Context Length | Variant | Download |
|---|---|---|---|---|
| salesforce/xgen-small-4B-base-r | 4B | 128k | Pre-trained | 🤗 Link |
| salesforce/xgen-small-4B-instruct-r | 4B | 128k | Post-trained | 🤗 Link |
| salesforce/xgen-small-9B-base-r | 9B | 128k | Pre-trained | 🤗 Link |
| salesforce/xgen-small-9B-instruct-r | 9B | 128k | Post-trained | 🤗 Link |
Usage
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "Salesforce/xgen-small-9B-instruct-r"
tokenizer = AutoTokenizer.from_pretrained(model_name)
device = "cuda" if torch.cuda.is_available() else "cpu"
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype="auto"
).to(device)
prompt = "What is Salesforce?"
messages = [{"role": "user", "content": prompt}]
inputs = tokenizer.apply_chat_template(
messages,
add_generation_prompt=True,
return_tensors="pt"
).to(model.device)
generated = model.generate(inputs, max_new_tokens=128)
output = tokenizer.decode(
generated[0],
skip_special_tokens=True,
)
print(output)
Evaluation
| Category | Task | Llama 3.1-8B | Granite 3.3-8B | Qwen2.5-7B | xGen-small 9B Instruct |
|---|---|---|---|---|---|
| General Knowledge & Reasoning | MMLU | 68.3 | 62.7 | 72.4 | 72.4 |
| General Knowledge & Reasoning | MMLU-Pro | 43.2 | 43.5 | 56.7 | 57.3 |
| Chat | Arena-Hard-v1.0 | 28.9 | 30.5 | 48.1 | 60.1 |
| Chat | MT-Bench | 8.25 | 8.57 | 8.56 | 8.90 |
| Math & Science | GPQA | 31.9 | 35.3 | 32.6 | 45.8 |
| Math & Science | GSM8K | 84.2 | 89.4 | 91.9 | 95.3 |
| Math & Science | MATH | 48.9 | 70.9 | 74.6 | 91.6 |
| Math & Science | AIME 2024 | 6.7 | 10.0 | 6.7 | 50.0 |
| Coding | HumanEval+ | 61.6 | 65.9 | 74.4 | 78.7 |
| Coding | MBPP+ | 55.3 | 60.3 | 68.8 | 63.8 |
| Coding | LiveCodeBench | 10.3 | 10.3 | 12.1 | 50.6 |
Citation
@misc{xgensmall,
title={xGen-small Technical Report},
author={Erik Nijkamp and Bo Pang and Egor Pakhomov and Akash Gokul and Jin Qu and Silvio Savarese and Yingbo Zhou and Caiming Xiong},
year={2025},
eprint={2505.06496},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2505.06496},
}
Ethical Considerations
This release is for research purposes only in support of an academic paper. Our models, datasets, and code are not specifically designed or evaluated for all downstream purposes. We strongly recommend users evaluate and address potential concerns related to accuracy, safety, and fairness before deploying this model. We encourage users to consider the common limitations of AI, comply with applicable laws, and leverage best practices when selecting use cases, particularly for high-risk scenarios where errors or misuse could significantly impact people's lives, rights, or safety. For further guidance on use cases, refer to our AUP and AI AUP.
Model Licenses
The models are being released under CC-BY-NC-4.0, Copyright © Salesforce, Inc. All Rights Reserved.