mesh-ops/K2-V2-base_final-GGUF

K2-V2

📚 Tech Report - 📝 Training Code - 🏢 Evaluation Code

🗂️ Pretraining Data: TxT360 - 🗂️ Midtraining Data: TxT360-Midas - 🗂️ SFT Data: TxT360-3efforts

K2-V2 is our most capable fully open model to date, and one of the strongest open-weight models in its class. It uses a 70B-parameter dense transformer architecture and represents the latest advancement in the LLM360 model family.

Beyond standard competencies such as factual knowledge and conversational ability, K2-V2 demonstrates strong long-context consistency, deep mathematical understanding, and robust reasoning skills. These capabilities serve as building blocks for sophisticated downstream applications, such as solving complex math problems and executing agentic workflows.

---

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("LLM360/K2-V2", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("LLM360/K2-V2")

prompt = "Explain why the derivative of sin(x) is cos(x)."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Using specific checkpoints

To make experimentation easier, we uploaded multiple checkpoints from each pretraining and midtraining stage as tagged releases. These checkpoints follow a consistent naming convention.

Checkpoints from different stages use the stage name as a prefix: base, mid_1, mid_2, mid_3, and mid_4. Final checkpoints are marked with the final suffix, while intermediate checkpoints are identified by their checkpoint numbers.

For example, the final checkpoint from midtraining stage 4 is named mid_4_final.

Quick links to all the final checkpoints for different stages:

stage	link
pretrain	https://huggingface.co/LLM360/K2-V2/tree/base_final
mid_1	https://huggingface.co/LLM360/K2-V2/tree/mid_1_final
mid_2	https://huggingface.co/LLM360/K2-V2/tree/mid_2_final
mid_3	https://huggingface.co/LLM360/K2-V2/tree/mid_3_final
mid_4	https://huggingface.co/LLM360/K2-V2/tree/mid_4_final

To use a specific checkpoint, use the revision argument:

model = AutoModelForCausalLM.from_pretrained("llm360/k2-v2", device_map="auto", revision="base_0720000")

---

Evaluation Summary

Below we report performance across general, reasoning, mathematical, and coding benchmarks. Scores for K2-V2 checkpoints (base → mid-4) demonstrate the impact of staged mid-training on reasoning quality.

Task / Model	base	mid-1	mid-2	mid-3	mid-4	Qwen2.5-72B	Llama3.0-70B	Llama3.1-70B	Olmo3-32B
General Tasks
MMLU	74.3	74.4	73.5	75.0	75.2	86.1	79.5	79.3	75.2
MMLU-Pro	43.7	46.8	48.1	59.8	57.0	58.1	52.8	53.8	49.6
BBH	68.4	79.8	81.1	82.2	83.2	86.3	82.2	82.1	77.6
HELLASWAG	87.8	86.9	86.6	86.6	86.0	87.6	88.0	85.0	84.8
WINOGRANDE	82.6	83.7	83.7	83.7	83.0	83.9	85.3	79.8	90.3
PIQA	84.2	84.0	83.3	82.9	83.1	83.5	84.6	84.3	85.6
TRUTHFULQA	54.0	54.9	55.1	55.8	53.9	60.5	45.6	49.7	54.9
Math & STEM Tasks
GPQA-DIAMOND	26.3	31.3	27.8	43.9	55.1	34.9	21.2	27.3	30.3
GSM8K	68.0	76.4	82.1	93.6	92.5	91.2	83.2	81.1	80.5
MATH	27.8	38.2	41.1	94.7	91.4	58.5	41.9	41.6	43.4
AIME 2025	0.0	17.6	25.1	53.2	46.9	1.7	0.1	0.2	14.7
ARC-CHALLENGE	64.9	66.4	66.4	66.0	66.3	72.4	69.2	64.9	65.4
Coding Tasks
MBPP	57.6	57.8	58.2	59.8	61.8	75.4	69.2	64.4	60.2
HUMANEVAL	50.0	51.2	53.7	54.3	54.3	54.3	42.1	50.6	36.0
Logic Puzzles
COUNTDOWN	1.3	53.3	53.1	35.9	75.6	6.0	1.0	0.5	23.2
KK-4 PEOPLE	4.8	44.9	68.0	64.5	92.9	26.1	4.2	7.6	42.4
KK-8 PEOPLE	0.5	23.2	41.3	51.6	82.8	5.7	1.1	1.3	13.0
ORDER-15 ITEMS	4.7	30.7	47.2	55.8	87.6	37.0	3.5	4.5	25.0
ORDER-30 ITEMS	0.0	0.3	3.0	34.1	40.3	0.7	0.2	0.1	0.6
Instruction Following
IFEVAL	17.4	26.2	28.5	34.5	26.7	40.3	15.1	17.4	13.2
Arabic
MMLU-Arabic	65.4	66.1	64.5	66.6	65.5	74.1	65.0	66.8	47.8

Below we report the evaluation results for K2-V2 after supervised fine-tuning (SFT). These variants correspond to three levels of reasoning effort (Low < Medium < High).

Metric / Model	K2 Low Dense · 70B	K2 Medium Dense · 70B	K2 High Dense · 70B	Olmo3 Think SFT Dense · 32B · No RL	Olmo3 Think Dense · 32B · RL	GLM-4.5 Air MoE · 106B A12B	MiniMax-M2 MoE · 230B A10B	Qwen3 235B MoE · 235B A22B · Reasoning	Qwen 2.5 72B Dense · 72B
LongBench V2	40.7	41.3	42.6	42.8	47.1	49.4	55.8	60.9	47.2
AIME25	27.3	62.0	80.2	68.3	73.3	81.3	75.8	88.8	15.2
HMMT25	19.0	45.6	71.4	43.3	50.83	73.3	63.5	84.2	9.79
GSM8K	92.4	92.0	94.8	96.1	95.7	96.1	95.4	93.5	85.8
Minerva	85.0	90.6	94.5	96.9	97.3	94.9	85.3	98.0	82.1
GPQA-D	48.5	60.6	69.3	58.0	59.8	75.3	76.2	80.7	50.5
MBPP	71.0	75.8	84.8	87.6	91.6	82.8	83.8	96.2	80.0
HumanEval	82.3	91.5	91.5	96.3	96.3	97.6	89.6	94.5	85.4
LCBv6	39.9	51.3	67.0	67.9	67.6	67.8	79.2	72.8	36.7

Please refer to our Tech Report for detailed evaluation results.

---

Datasets & Mixtures

K2-V2 training is organized into three stages, each using a transparent, publicly released mixture:

Pretraining Mix

• Large-scale natural text corpus spanning web content, books, code, and multilingual sources
• Mixture designed for stable scaling and broad general-knowledge coverage
• ~12T tokens

Mid-Training Mix

• TxT360-Midas: reasoning-oriented + long-context extensions
• Domain-focused sources: math, programming, scientific literature
• Synthetic expansions where natural data is scarce

SFT Mix

• Check out https://huggingface.co/LLM360/K2-V2-Instruct

All mixtures, filtering rules, and data sources are fully released for reproducibility.

Please refer to our Tech Report for detailed datasets and mixtures information.

---

Model Description

• Model type: K2-V2 follows a standard decoder-only transformer with grouped-query attention and RMSNorm.
• Training stage: Pre-training
• Language(s) (NLP): English
• License: Apache 2.0

Model Hyperparameter	Value
Total Parameters	70B
Hidden Size	8,192
Intermediate Size (FFN)	28,672
Number of Attention Heads	64
Number of Layers	80
RMSNorm ɛ	1e-5
Pre-training Seq Length	8,192
Max Mid-training Seq Length	524,288
Vocab Size	250,000

---

Intended Use

K2-V2 is designed for:

• research on large language models and reasoning
• downstream fine-tuning (e.g., instruction following, agents, domain models)
• experimentation with long-context architectures
• open, transparent benchmarking of LLM scaling

K2-V2 is not instruction-tuned. For aligned conversational use, please see K2-V2-Instruct.

---

Limitations

• May generate incorrect or hallucinated content, especially when asked about facts not seen during training
• Not optimized for safety, moderation, or refusal behavior (base model)
• Long-context performance depends on prompt quality and retrieval structure
• Primarily trained on English; multilingual capabilities are limited
• Inference cost is high due to the 70B parameter size

---

Citation

If you use K2-V2 in your research, please cite the following:

@misc{k2team2025k2v2360openreasoningenhancedllm,
      title={K2-V2: A 360-Open, Reasoning-Enhanced LLM}, 
      author={K2 Team and Zhengzhong Liu and Liping Tang and Linghao Jin and Haonan Li and Nikhil Ranjan and Desai Fan and Shaurya Rohatgi and Richard Fan and Omkar Pangarkar and Huijuan Wang and Zhoujun Cheng and Suqi Sun and Seungwook Han and Bowen Tan and Gurpreet Gosal and Xudong Han and Varad Pimpalkhute and Shibo Hao and Ming Shan Hee and Joel Hestness and Haolong Jia and Liqun Ma and Aaryamonvikram Singh and Daria Soboleva and Natalia Vassilieva and Renxi Wang and Yingquan Wu and Yuekai Sun and Taylor Killian and Alexander Moreno and John Maggs and Hector Ren and Guowei He and Hongyi Wang and Xuezhe Ma and Yuqi Wang and Mikhail Yurochkin and Eric P. Xing},
      year={2025},
      eprint={2512.06201},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2512.06201}, 
}

---