Ring-lite-2507

Introduction

We present Ring-lite-2507, an upgraded version of our previously released lightweight reasoning model, Ring-lite. Building upon 16.8B Mixture-of-Experts (MoE)-based large language model with 2.75B activated parameters, Ring-lite-2507 further pushes its reasoning ability to an advanced level, meanwhile, it demonstrates superior performance on a comprehensive range of LLM benchmarks, including general text understanding, alignment, coding, logical and agentic tasks. Thanks to our innovative and robust reinforcement learning training pipeline, Ring-lite-2507 distinguished itself from latest public dense models under 10B parameters by showing competitive performance across various tasks while activating only 1/3 of their parameter size.

Model Downloads

Model	#Total Params	#Activated Params	Context Length	Download
Ring-lite-2507	16.8B	2.75B	128K	🤗 HuggingFace
Ring-lite	16.8B	2.75B	128K	🤗 HuggingFace

Evaluation

For a comprehensive evaluation of the quality of our reasoning models, we implemented automatic benchmarks to assess their performance including math, code and science.

To compare the performance of Ring-lite-2507 and Ring-lite, we evaluate the two models on a broader range of reasoning and general-purpose benchmarks, including knowledge understanding, math, coding, reasoning & agentic and alignment.

Knowledge Understanding

Benchmark	Ring-lite-2507	Ring-lite-2506	Qwen3-8B-Thinking
MMLU-Pro (EM)	72.50	63.44	72.56
GPQA-Diamond (Pass@1)	69.35	63.51	62.00
SuperGPQA (EM)	40.05	13.97	40.36
Phybench (Pass@1)	28.51	29.19	22.14

Math

Benchmark	Ring-lite-2507	Ring-lite-2506	Qwen3-8B-Thinking
MATH-500 (Pass@1)	97.95	96.80	97.30
CNMO 2024 (Pass@1)	75.09	77.26	74.57
AIME 2024 (Pass@1)	79.79	79.00	74.90
AIME 2025 (Pass@1)	72.92	69.50	67.19
LiveMathBench (Pass@1)	83.37	85.08	81.90
TheoremQA (Pass@1)	70.00	70.19	68.81
OlympiadBench (math) (Pass@1)	80.64	82.86	80.20

Coding

Benchmark	Ring-lite-2507	Ring-lite-2506	Qwen3-8B-Thinking
LiveCodeBench(2408-2505) (Pass@1)	60.35	59.53	55.12
Codeforces(Percentile) (Pass@1)	1830	1673	1580
Codeforces(Rating)	92.16	88.00	79.44

Reasoning & Agentic

Benchmark	Ring-lite-2507	Ring-lite-2506	Qwen3-8B-Thinking
DROP (zero-shot F1)	89.27	60.21	87.13
BBH (EM)	88.65	50.84	87.30
ARCPrize (Pass@1)	19.00	3.12	3.88
MuSR (EM)	77.19	66.77	76.92
BFCL_Live (Pass@1)	74.81	66.76	75.99

Alignment

Benchmark	Ring-lite-2507	Ring-lite-2506	Qwen3-8B-Thinking
IFEval (Prompt Strict)	84.66	54.34	85.40
AlignBench v1.1(gpt-4.1)	80.90	69.60	74.70
FoFo (gpt-4-turbo)	85.02	67.81	81.93
ArenaHard (gpt-4.1)	88.85	56.12	86.14

Blog

More details are reported in our blog.

Quickstart

🤗 Hugging Face Transformers

Here is a code snippet to show you how to use the chat model with transformers:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "inclusionAI/Ring-lite-2507"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = "Give me a short introduction to large language models."
messages = [
    {"role": "system", "content": "You are Ring, an assistant created by inclusionAI"},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=8192
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

Deployment

Please refer to GitHub

License

This code repository is licensed under the MIT License.

Citation

@misc{ringteam2025ringlitescalablereasoningc3postabilized,
      title={Ring-lite: Scalable Reasoning via C3PO-Stabilized Reinforcement Learning for LLMs}, 
      author={Ling Team},
      year={2025},
      eprint={2506.14731},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2506.14731}, 
}

inclusionAI
/

Ring-lite-2507