IFM/K2-Think-V2

1---2base_model:3- LLM360/K2-V2-Instruct4language:5- en6library_name: transformers7license: apache-2.08pipeline_tag: text-generation9---10 11# K2 Think V2: A Fully-Sovereign Reasoning Model12 13📚 [Blog](https://mbzuai.ac.ae/news/k2-think-v2-a-fully-sovereign-reasoning-model) - 📝 [Code](https://github.com/LLM360/Reasoning360) - 🏢 [Project Page](https://k2think.ai)14 15<center><img src="banner.png" alt="k2-think-banner"/></center>16 17<br>18 19K2 Think V2 is a 70 billion parameter open-weights general reasoning model with strong performance in competitive mathematical problem solving built on-top of [K2-V2-Instruct](https://huggingface.co/LLM360/K2-V2-Instruct), comprising a fully sovereign reasoning model.20 21# Quickstart22 23### Serving configurations24 25We use the following serving configurations:26 27| Setting | Value |28| - | - |29| Temperature | 1.0 |30| Top-p | 1.0 |31| Top-k | -1 |32| Context Length | 131072 |33| Context Length Extension | 2x using YaRN |34| Chat Template | Default provided in `chat_template.jinja` |35 36The provided chat template sets the reasoning effort to `high`37 38### Serving commands39 40To serve on VLLM:41 42```43vllm serve LLM360/K2-Think-V2 --tensor-parallel-size 8 --port 800044```45 46### Transformers47You can use `K2 Think V2` 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.48 49The chat template is directly inherited from K2-V2-Instruct, with the default `reasoning_effort` set to `"high"`. The other levels of reasoning effort (`"low"` and `"medium"`) are still available but have not been tested or evaluated. As such, the model's behavior under such settings is not assured to maintain reported performance. 50 51```python52from transformers import pipeline53import torch54 55model_id =  "LLM360/K2-Think-V2"56 57pipe = pipeline(58    "text-generation",59    model=model_id,60    torch_dtype="auto",61    device_map="auto",62)63 64messages = [65    {"role": "user", "content": "what is the next prime number after 2600?"},66]67 68outputs = pipe(69    messages,70    max_new_tokens=131072,71)72print(outputs[0]["generated_text"][-1])73```74 75If you cannot use `tokenizer.apply_chat_template`, you may also pass in these arguments using `extra_body` and `chat_template_kwargs`:76 77```78from openai import OpenAI79 80client = OpenAI(81    base_url="http://localhost:8000/v1",82    api_key="key"83)84 85completion = client.chat.completions.create(86    model="LLM360/K2-Think-V2",87    messages = [88        {"role": "system", "content": "You are K2-Think, a helpful assistant created by Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) Institute of Foundation Models (IFM)."},89        {"role": "user", "content": "Solve the 24 game [2, 3, 5, 6]"}90    ],91    extra_body={92        "chat_template_kwargs": {"reasoning_effort": "high"},93    },94)95```96 97---98 99# Evaluation & Performance100A more complete summary of evaluation results are reported in our [Blog](https://mbzuai.ac.ae/news/k2-think-v2-a-fully-sovereign-reasoning-model)101 102## Benchmarks (pass\@1, average over 16 runs)103 104| Domain  | Benchmark             |  K2 Think V2 |105| ------- | --------------------  | -----------: |106| Math    | AIME 2025             |        90.42 |107| Math    | HMMT 2025             |        84.79 |108| Code    | SciCode               |        33.00 |109| Science | GPQA-Diamond          |        72.98 |110| Science | Humanity's Last Exam  |        9.5 |111 112## Safety Evaluation113 114Aggregated across four safety dimensions (**Safety-4**):115 116K2 Think V2 establishes a robust safety baseline while effectively resolving the "alignment tax" of [previous K2 Think](https://huggingface.co/LLM360/K2-Think) releases. Despite strong overall safety performance, there are still opportunities to improve the model with regard to handling sensitive personal information.117 118| Safety Surface                  | Macro-Avg | Risk Level |119| ------------------------------- | --------: | ---------- |120| Content & Public Safety         |     98.20 |    Low     |121| Truthfulness & Reliability      |     97.98 |    Low     |122| Societal Alignment              |     97.25 |    Low     |123| Data & Infrastructure           |     83.00 |  Critical  |124 125---126 127# Terms of Use128 129We have employed various techniques to reduce bias, harmful outputs, and other risks in the model. While these efforts help improve safety and reliability, the model, like all Large Language Models, may still generate inaccurate, misleading, biased, or otherwise undesirable content. By downloading, using, or interacting with this model, you acknowledge these limitations and agree to the following:130 1311. **Prohibited Uses**  132   - You may **not** use this model for any **illegal, unlawful, or harmful activities**, including but not limited to fraud, abuse, harassment, privacy violations, or the creation/dissemination of malicious content.  133 1342. **User Responsibility**  135   - You are solely responsible for how you use the model and for any outcomes that result from its use.  136   - The authors and institutions involved in releasing this model do **not** accept liability for any consequences arising from its use.  137 1383. **No Warranty**  139   - The model is provided **“as is” without any warranties or guarantees**.  140---141 142# Citation143If you use K2 Think V2 in your research, please use the following citation:144 145```bibtex146@misc{k2think2026k2think0126,147      title={K2 {T}hink {V}2: A {F}ully-{S}overeign {R}easoning {M}odel}, 148      author={K2 Think Team and Taylor W. Killian and Varad Pimpalkhute and Richard Fan and Haonan Li and Chengqian Gao and Ming Shan Hee and Xudong Han and John Maggs and Guowei He and Zhengzhong Liu and Eric P. Xing},149      year={2026},150      url={https://mbzuai.ac.ae/news/k2-think-v2-a-fully-sovereign-reasoning-model}, 151}152```