---
language:
- en
license: apache-2.0
pipeline_tag: text-generation
tags:
- text-generation
- agent
- tool-use
- long-context
library_name: transformers
---
# LIMI: Less is More for Agency
[](https://arxiv.org/pdf/2509.17567)
[](https://github.com/GAIR-NLP/LIMI)
[](https://huggingface.co/datasets/GAIR/LIMI)
---
To learn more about LIMI, feel free to explore our documentation and resources. Our release consists of the following sections:
- **Model Zoo && Quick Start**: Basic usage and demonstrations with Transformers, vLLM, and SGLang for LIMI and LIMI-Air;
- **Evaluation**: Comprehensive evaluation suite with metrics for agentic capabilities assessment;
- **Prompting**: Usage of LIMI with frameworks for agentic applications, tool use, and reasoning tasks.
## Overview
LIMI is an agentic model fine‑tuned from [GLM‑4.5](https://huggingface.co/zai-org/GLM-4.5) using compact, high‑quality data to emphasize:
- Targeted capabilities: tool use, multi‑turn correction, spec compliance
- Long‑context trajectory with tokenizer‑filtered samples
- OpenAI‑style `messages` with optional function/tool calls
## Model Details
- Base model: `zai-org/GLM-4.5`
- Training framework: slime
- Training data: curated conversations from [GAIR/LIMI](https://huggingface.co/datasets/GAIR/LIMI)
## Performance
### SFT with LIMI Dataset on Dense Models
Our LIMI dataset significantly enhances dense models (Qwen3 series) on both in-domain and out-of-domain benchmarks:
The figure above demonstrates the effectiveness of our training approach:
- **Left (AgencyBench)**: Substantial improvements on in-domain agentic tasks, with Qwen3-4B (4.6% → 8.6%), Qwen3-8B (7.3% → 10.6%), and Qwen3-32B (8.4% → 20.5%).
- **Right (Out-of-Domain)**: Strong generalization to unseen benchmarks while maintaining performance, with Qwen3-4B (28.3% → 28.9%), Qwen3-8B (31.2% → 32.0%), and Qwen3-32B (35.2% → 37.1%).
### LIMI Models on AgencyBench
Our models achieve state-of-the-art performance across multiple agentic evaluation tasks:
| Model | FTFC (↑) | RC@3 (↑) | SR@3 (↑) | Avg. |
|-------|----------|----------|----------|-----------------|
| GLM-4.5-Air | 15.0 | 16.1 | 20.0 | 17.0 |
| GLM-4.5 | 37.8 | 50.0 | 47.4 | 45.1 |
|GLM-4.5-Code| 48.0 | 48.0|47.5| 47.8|
| **LIMI-Air** | **35.4** | **34.3** | **33.1** | **34.3** |
| **LIMI** | **71.7** | **74.2** | **74.6** | **73.5** |
For detailed benchmark results, experimental setup, and comprehensive comparisons, please refer to our [paper](https://arxiv.org/pdf/2509.17567).
## Model Zoo
Our LIMO model is available on Hugging Face 🤗:
| Model | Backbone | Size | Link |
|---|---|---|---|
| LIMI | [GLM‑4.5](https://huggingface.co/zai-org/GLM-4.5) | 353B | https://huggingface.co/GAIR/LIMI |
| LIMI‑Air | [GLM‑4.5‑Air](https://huggingface.co/zai-org/GLM-4.5-Air) | 107B | https://huggingface.co/GAIR/LIMI-Air |
## Datasets
We release our datasets through Hugging Face 🤗:
- Name: `GAIR/LIMI`
- Summary: curated agentic SFT data (OpenAI `messages`, optional `tools`, normalized tool‑call arguments); current release contains ~78 high‑quality samples.
- Link: https://huggingface.co/datasets/GAIR/LIMI
## Quick Start
Start with HF Transformers
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model = AutoModelForCausalLM.from_pretrained(
"GAIR/LIMI", torch_dtype="auto", device_map="auto", trust_remote_code=True
)
tok = AutoTokenizer.from_pretrained("GAIR/LIMI", trust_remote_code=True)
messages = [
{"role": "system", "content": "You are a helpful assistant tasked with discovering mathematical function structures for scientific systems."},
{"role": "user", "content": "Modify the equation.py function, considering the physical meaning and relationships of the inputs."}
]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(text, return_tensors="pt").to(model.device)
out = model.generate(
**inputs,
max_new_tokens=4096,
temperature=0.6,
top_p=0.95,
do_sample=True,
)
print(tok.decode(out[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True))
```
Start with VLLM
```python
from vllm import LLM, SamplingParams
from transformers import AutoTokenizer
llm = LLM(model="GAIR/LIMI", trust_remote_code=True)
tok = AutoTokenizer.from_pretrained("GAIR/LIMI", trust_remote_code=True)
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
out = llm.generate(text, SamplingParams(temperature=0.6, max_tokens=4096, top_p=0.95))
print(out[0].outputs[0].text)
```
## Prompting
- Messages follow OpenAI chat format; include a grounding system message when helpful.
- Example:
```json
[
{"role": "system", "content": "You are a helpful assistant tasked with discovering mathematical function structures for scientific systems."},
{"role": "user", "content": "Modify the equation.py function, considering the physical meaning and relationships of the inputs."}
]
```
## Evaluation
- We report FTFC (First‑Turn Functional Completeness), SR@R (Success Rate at R), and RC@R (Remaining Chances at R) with R=3.
- See the paper for experimental protocol and scores.
## Limitations
- May produce incorrect tool arguments or overfit to frequent schemas
- Not safety‑filtered for sensitive domains; use with guardrails and oversight
## License
- Inherits base model (GLM‑4.5) terms; verify upstream license before deployment
## Citation
```bibtex
@misc{xiao2025limiagency,
title={LIMI: Less is More for Agency},
author={Yang Xiao and Mohan Jiang and Jie Sun and Keyu Li and Jifan Lin and Yumin Zhuang and Ji Zeng and Shijie Xia and Qishuo Hua and Xuefeng Li and Xiaojie Cai and Tongyu Wang and Yue Zhang and Liming Liu and Xia Wu and Jinlong Hou and Yuan Cheng and Wenjie Li and Xiang Wang and Dequan Wang and Pengfei Liu},
year={2025},
eprint={2509.17567},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2509.17567},
}
```
