language:
- en
license: apache-2.0
pipeline_tag: text-generation
library_name: transformers
base_model: Qwen/Qwen2.5-32B
Try K2-Think · Tech Report · Code
K2-Think is a 32 billion parameter open-weights general reasoning model with strong performance in competitive mathematical problem solving. Built on a Qwen2.5-32B base, K2-Think combines long CoT SFT, RL with verifiable rewards, and a test-time scaling scaffold to match or exceed much larger models on public math benchmarks while keeping latency low.
Highlights
- Math specialist at 32B: State-of-the-art results among open models on AIME-style olympiad math and other hard math sets.
- Fast generation: ~2,000 tokens/sec on our Cerebras WSE deployment; ~10× faster than typical H100/H200 setups in our tests.
- Token-efficient reasoning: Planning reduces average response length by up to ~14% at equal or higher accuracy.
Quickstart
Transformers
You can use K2-Think with Transformers. If you use transformers.pipeline, it will apply the chat template automatically. If you use model.generate directly, you need to apply the chat template mannually.
from transformers import pipeline
import torch
model_id = "LLM360/K2-Think"
pipe = pipeline(
"text-generation",
model=model_id,
torch_dtype="auto",
device_map="auto",
)
messages = [
{"role": "user", "content": "what is the next prime number after 2600?"},
]
outputs = pipe(
messages,
max_new_tokens=32768,
)
print(outputs[0]["generated_text"][-1])
Evaluation & Performance
Detailed evaluation results are reported in out Tech Report
Benchmarks (pass@1, average over 16 runs)
| Domain | Benchmark | K2-Think |
|---|---|---|
| Math | AIME 2024 | 90.83 |
| Math | AIME 2025 | 81.24 |
| Math | HMMT 2025 | 73.75 |
| Math | OMNI-Math-HARD | 60.73 |
| Code | LiveCodeBench v5 | 63.97 |
| Science | GPQA-Diamond | 71.08 |
Inference Speed
We deploy K2-THINK on Cerebras Wafer-Scale Engine (WSE) systems, leveraging the world’s largest processor and speculative decoding to achieve unprecedented inference speeds for our 32B reasoning system.
| Platform | Throughput (tokens/sec) | Example: 32k-token response (time) |
|---|---|---|
| Cerebras WSE (our deployment) | ~2,000 | ~16 s |
| Typical H100/H200 GPU setup | ~200 | ~160 s |
Token Efficiency
K2-Think's Plan-Before-You-Think methodology combined with Best-of-N sampling produces more concise reasoning chains while maintaining or improving accuracy. Our test-time scaffold reduces average response length by up to 14% across mathematical benchmarks.
Token reduction per completed answer (SFT+RL checkpoint vs K2-Think):
| Domain | Benchmark | SFT+RL Checkpoint | K2-Think | Δ |
|---|---|---|---|---|
| Math | AIME24 | 23,324 | 20,058 | −14.0% |
| Math | AIME25 | 25,869 | 24,218 | −6.38% |
| Math | HMMT25 | 31,475 | 26,977 | −14.3% |
| Math | OMNI-Math-HARD | 35,266 | 30,032 | −14.0% |
| Code | LiveCodeBench | 13,552 | 12,166 | −10.2% |
| Science | GPQA-Diamond | 15,271 | 14,661 | −3.99% |
Safety Evaluation
Aggregated across four safety dimensions (Safety-4):
| Aspect | Macro-Avg |
|---|---|
| High-Risk Content Refusal | 0.830 |
| Conversational Robustness | 0.890 |
| Cybersecurity & Data Protection | 0.560 |
| Jailbreak Resistance | 0.705 |
| Safety-4 Macro (avg) | 0.746 |
Citation
@techreport{k2think2025,
title = {K2-Think: A Parameter-Efficient Reasoning System},
author = {Zhoujun Cheng* and Richard Fan* and Shibo Hao* and Taylor W. Killian* and Haonan Li* and Suqi Sun* and Hector Ren and Alexander Moreno and Daqian Zhang and Tianjun Zhong and Yuxin Xiong and Yuanzhe Hu and Yutao Xie and Xudong Han and Yuqi Wang and Varad Pimpalkhute and Yonghao Zhuang and Aaryamonvikram Singh and Xuezhi Liang and Anze Xie and Jianshu She and Desai Fan and Chengqian Gao and Liqun Ma and Mikhail Yurochkin and John Maggs and Xuezhe Ma and Guowei He and Zhiting Hu and Zhengzhong Liu and Eric P. Xing},
year = {2025},
institution = {Institute of Foundation Models, Mohamed bin Zayed University of Artificial Intelligence},
url = {https://k2think.ai}
}