Kimi Releases K2 Pondering — Open-Provide “Pondering” Model That Boosts Agent And Reasoning Expertise

Kimi proper now launched Kimi K2 Pondering, the company’s most succesful open-source “pondering” model thus far. Constructed throughout the “model-as-agent” concept, K2 Pondering natively combines prolonged multi-step reasoning with in depth machine use — enabling brokers which will “suppose whereas using devices.”

What it does

Kimi says K2 Pondering can autonomously run as a lot as 300 tool-call cycles in a single session and keep prolonged, safe multi-turn reasoning chains. That capability is powered by the group’s latest Check out-Time Scaling strategies, which delay every the number of reasoning tokens and tool-call iterations at inference time to reinforce agentic and reasoning effectivity.

Benchmarks and capabilities

K2 Pondering achieves state-of-the-art (SOTA) outcomes all through quite a few agent and reasoning benchmarks:

• Humanity’s Ultimate Examination (an entire closed-book tutorial check out spanning 100+ disciplines): 44.9% (SOTA when devices are permitted).
• BrowseComp (OpenAI’s benchmark for web-browsing brokers): 60.2% (new SOTA; human frequent is ~29.2%).
• SEAL-0 and completely different sophisticated information-gathering/reasoning exams: SOTA-level effectivity.
  Kimi highlights good factors in agentic search, agentic programming, inventive writing and fundamental multi-step reasoning. Occasion walkthroughs current the model chaining iterative search → browse → code → reasoning loops to decompose open-ended points into actionable subtasks and produce verified options.

Agentic coding and inventive duties

K2 Pondering improves coding effectivity on multilingual software-engineering benchmarks (SWE-Multilingual, SWE-bench, Terminal duties). The model is finest at front-end duties (HTML/React/components) and may perform inside software program program brokers to deal with multi-step development workflows — as an illustration, assembling a functioning Phrase-style editor or producing voxel-art creations.
Creative and evaluation capabilities are moreover stronger: the model produces additional coherent long-form inventive writing, deeper tutorial analysis, and further empathetic, wise responses to private or emotional queries.

Effectivity: native INT4 quantization

To chop again latency and GPU memory utilization all through prolonged reasoning runs, Kimi utilized quantization-aware teaching and weight-only INT4 quantization for MoE components. The consequence: native INT4 inference help that roughly doubles know-how tempo and improves compatibility with dwelling accelerator chips. Kimi notes that all reported benchmark scores have been obtained beneath INT4 precision.

Availability

K2 Pondering is already dwell on kimi.com and throughout the latest Kimi cell app beneath the standard chat mode. The underlying model may even alternate the underside model in Kimi’s Agent mode in a forthcoming substitute to permit full multi-turn pondering and energy use.
Builders can entry the model by the use of the Kimi Open Platform or receive it from public model hubs equal to Hugging Face and ModelScope for self-hosting. The platform helps 256K context.

Notes on deployed experience

To keep up the frequent chat experience lightweight, Kimi deploys a restricted machine set and fewer tool-call rounds on kimi.com and throughout the app. In consequence, on-site chat couldn’t match benchmark scores; the entire agentic capabilities will develop to be seen when the Agent mode (“OK Computer”) is updated to K2 Pondering.

provide：Kimi