K2 Think 32B rivals reasoning models 20 times its size

K2 Think achieves 2,000 tokens/sec for real-time reasoning at scale with Cerebras

#UAE #Reasoning - The Institute of Foundation Models lab established by Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and Abu Dhabi’s AI powerhouse G42 have released K2 Think, an open-source advanced reasoning system engineered for performance, transparency, and efficiency. According to the university’s benchmark tests, the new 32 billion parameter reasoning model outperforms flagship reasoning models 20 times larger. The breakthrough model ranks among the industry's top reasoning systems, leading all open-source models in mathematics performance across AIME '24/'25, HMMT '25, and OMNI-Math-HARD benchmarks, whilst offering full transparency including training data, parameter weights, software code for deployment, and test-time optimisation unlike most "open" AI models that only release weights.

SO WHAT? - This launch of K2 Think positions the UAE among global leaders in efficient AI development, proving that advanced reasoning capabilities don't necessarily require massive computational resources. The model release not only marks a milestone in MBZUAI’s model research, but could prove to be an important step toward frontier-class reasoning that the global AI community can study, use, and build on. With both open source data released and an online chat app with super-fast inference running on Cerebras Wafer-Scale Engine (WSE) systems, K2 Think is bound to cause a stir and inspire debate.

Here are the key details about the K2 Think release:

• The Institute of Foundation Models lab established by Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and Abu Dhabi’s AI powerhouse G42 have released K2 Think, an open-source advanced reasoning system engineered for performance, transparency, and efficiency.

• K2 Think employs six pillars of innovation:

  • Long chain-of-thought supervised fine-tuning

  • Reinforcement learning with verifiable rewards

  • Agentic planning for complex problem decomposition

  • Test-time scaling for adaptability

  • Speculative decoding optimized for throughput

  • Inference-ready hardware integration

• According to MBZUAI, K2 Think leads all open-source models in math performance on AIME ’24/’25, HMMT ’25, and OMNI-Math-HARD.

• The K2 Think release includes open-source training data, code, and optimisation recipes for full reproducibility, extensibility, and trust.

• A chat version of K2 Think been made available on Cerebras Systems' wafer-scale, inference-optimised compute platform, achieving unprecedented throughput of 2000 tokens per second with speculative decoding optimised for Cerebras hardware.

• The model builds on a growing family of UAE-developed open-source models including Jais (Arabic LLM), NANDA (Hindi), SHERKALA (Kazakh), and extends the legacy of K2-65B released in 2024.

• K2 Think represents the culmination of close public-private partnership between MBZUAI's academic research capabilities and G42's technological infrastructure and commercial expertise in Abu Dhabi's AI ecosystem.

ZOOM OUT - The K2 Think model developed by the Institute of Foundation Models has been released via the LLM360 community established by MZUAI, Californian AI supercomputer developer Cerebras Systems, and generative AI company Petuum. The community provides the global development community with code, data and insights to help them build tomorrow’s AI models. LLM360 addresses a critical gap in open-source AI development by providing complete transparency into LLM training processes, including code, data sequences, and detailed methodologies that traditionally remain proprietary. The community’s mission is to democratise AI research through open standards and knowledge transfer has positioned MBZUAI as a leader in transparent, accessible AI model development.

[Written and edited with the assistance of AI]

LINKS

• K2 Think chat (K2 website)

• K2 Think Data (Hugging Face)

• K2 Think technical report PDF (K2 website)

Read more about MBZUAI open source initiatives:

• UAE President backs UAE-made AI reasoning model (Middle East AI News)

• LLM360 project empowers pre-trainers (Middle East AI News)

• Powerful open-source K2-65B LLM costs 35% less to train (Middle East AI News)

• New framework for open-source LLMs (Middle East AI News)