How will we design AI methods that may plan, motive, and act over lengthy sequences of choices with out fixed human steerage? Moonshot AI has launched Kimi K2 Considering, an open supply considering agent mannequin that exposes the complete reasoning stream of the Kimi K2 Combination of Specialists structure. It targets workloads that want deep reasoning, lengthy horizon device use, and secure agent conduct throughout many steps.
What’s Kimi K2 Considering?
Kimi K2 Considering is described as the newest, most succesful model of Moonshot’s open supply considering mannequin. It’s constructed as a considering agent that causes step-by-step and dynamically invokes instruments throughout inference. The mannequin is designed to interleave chain of thought with perform calls so it will probably learn, suppose, name a device, suppose once more, and repeat for a whole bunch of steps.
The mannequin units a brand new state-of-the-art on Humanity’s Final Examination and BrowseComp, whereas sustaining coherent conduct throughout about 200 to 300 sequential device calls with out human interference.
On the similar time, K2 Considering is launched as an open weights mannequin with a 256K token context window and native INT4 inference, which reduces latency and GPU reminiscence utilization whereas preserving benchmark efficiency.
K2 Considering is already reside on kimi.com in chat mode and is accessible by means of the Moonshot platform API, with a devoted agentic mode deliberate to show the complete device utilizing conduct.
Structure, MoE design, and context size
Kimi K2 Considering inherits the Kimi K2 Combination of Specialists design. The mannequin makes use of a MoE structure with 1T complete parameters and 32B activated parameters per token. It has 61 layers together with 1 dense layer, 384 consultants with 8 consultants chosen per token, 1 shared knowledgeable, 64 consideration heads, and an consideration hidden dimension of 7168. The MoE hidden dimension is 2048 per knowledgeable.
The vocabulary dimension is 160K tokens and the context size is 256K. The eye mechanism is Multi head Latent Consideration, and the activation perform is SwiGLU.
Take a look at time scaling and lengthy horizon considering
Kimi K2 Considering is explicitly optimized for check time scaling. The mannequin is educated to develop its reasoning size and power name depth when going through more durable duties, relatively than counting on a hard and fast brief chain of thought.
On Humanity’s Final Examination within the no instruments setting, K2 Considering scores 23.9. With instruments, the rating rises to 44.9, and within the heavy setting it reaches 51.0. On AIME25 with Python, it reviews 99.1, and on HMMT25 with Python it reviews 95.1. On IMO AnswerBench it scores 78.6, and on GPQA it scores 84.5.
The testing protocol caps considering token budgets at 96K for HLE, AIME25, HMMT25, and GPQA. It makes use of 128K considering tokens for IMO AnswerBench, LiveCodeBench, and OJ Bench, and 32K completion tokens for Longform Writing. On HLE, the utmost step restrict is 120 with a 48K reasoning finances per step. On agentic search duties, the restrict is 300 steps with a 24K reasoning finances per step.
Benchmarks in agentic search and coding
On agentic search duties with instruments, K2 Considering reviews 60.2 on BrowseComp, 62.3 on BrowseComp ZH, 56.3 on Seal 0, 47.4 on FinSearchComp T3, and 87.0 on Frames.
On basic information benchmarks, it reviews 84.6 on MMLU Professional, 94.4 on MMLU Redux, 73.8 on Longform Writing, and 58.0 on HealthBench.
For coding, K2 Considering achieves 71.3 on SWE bench Verified with instruments, 61.1 on SWE bench Multilingual with instruments, 41.9 on Multi SWE bench with instruments, 44.8 on SciCode, 83.1 on LiveCodeBenchV6, 48.7 on OJ Bench within the C plus plus setting, and 47.1 on Terminal Bench with simulated instruments.
Moonshot crew additionally defines a Heavy Mode that runs eight trajectories in parallel, then aggregates them to provide a closing reply. That is utilized in some reasoning benchmarks to squeeze out further accuracy from the identical base mannequin.
Native INT4 quantization and deployment
K2 Considering is educated as a local INT4 mannequin. The analysis crew applies Quantization Conscious Coaching through the submit coaching stage and makes use of INT4 weight solely quantization on the MoE elements. This helps INT4 inference with roughly a 2x technology velocity enchancment in low latency mode whereas sustaining state-of-the-art efficiency. All reported benchmark scores are obtained beneath INT4 precision.
The checkpoints are saved in compressed tensors format and may be unpacked to increased precision codecs corresponding to FP8 or BF16 utilizing the official compressed tensors instruments. Really useful inference engines embrace vLLM, SGLang, and KTransformers.
Key Takeaways
- Kimi K2 Considering is an open weights considering agent that extends the Kimi K2 Combination of Specialists structure with express lengthy horizon reasoning and power use, not simply brief chat fashion responses.
- The mannequin makes use of a trillion parameter MoE design with about tens of billions of lively parameters per token, a 256K context window, and is educated as a local INT4 mannequin with Quantization Conscious Coaching, which provides about 2x quicker inference whereas protecting benchmark efficiency secure.
- K2 Considering is optimized for check time scaling, it will probably perform a whole bunch of sequential device calls in a single job and is evaluated beneath massive considering token budgets and strict step caps, which is vital if you attempt to reproduce its reasoning and agentic outcomes.
- On public benchmarks, it leads or is aggressive on reasoning, agentic search, and coding duties corresponding to HLE with instruments, BrowseComp, and SWE bench Verified with instruments, displaying that the considering oriented variant delivers clear features over the bottom non considering K2 mannequin.
Kimi K2 Considering is a powerful sign that check time scaling is now a firstclass design goal for open supply reasoning fashions. Moonshot AI is just not solely exposing a 1T parameter Combination of Specialists system with 32B lively parameters and 256K context window, it’s doing so with native INT4 quantization, Quantization Conscious Coaching, and power orchestration that runs for a whole bunch of steps in manufacturing like settings. Total, Kimi K2 Considering exhibits that open weights reasoning brokers with lengthy horizon planning and power use have gotten sensible infrastructure, not simply analysis demos.
Take a look at the Model Weights and Technical Details. Be happy to take a look at our GitHub Page for Tutorials, Codes and Notebooks. Additionally, be at liberty to observe us on Twitter and don’t overlook to affix our 100k+ ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
Asif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is dedicated to harnessing the potential of Synthetic Intelligence for social good. His most up-to-date endeavor is the launch of an Synthetic Intelligence Media Platform, Marktechpost, which stands out for its in-depth protection of machine studying and deep studying information that’s each technically sound and simply comprehensible by a large viewers. The platform boasts of over 2 million month-to-month views, illustrating its reputation amongst audiences.
