Kimi K2 Thinking is built on a Mixture-of-Experts (MoE) architecture with 1 trillion total parameters and 32 billion active parameters per token. The model employs 384 experts, dynamically selecting 8 per token plus 1 shared expert, enabling high specialization while maintaining computational efficiency. Its architecture comprises 61 layers and incorporates Multi-Head Latent Attention (MLA), which optimizes long-context processing by significantly reducing KV-cache overhead. A key distinction from conventional MoE models is the use of the MuonClip optimizer during pretraining, which achieved zero training instability even at the trillion-parameter scale. A unique feature of K2 Thinking is its built-in INT4 quantization, implemented via Quantization-Aware Training (QAT) during the fine-tuning phase. Quantization is applied exclusively to the MoE components, doubling generation speed with no perceptible loss in quality.
K2 Thinking is trained to interleave chain-of-thought reasoning with function calls, enabling it to autonomously execute research, coding, and writing workflows spanning hundreds of steps. The model sustains goal-directed behavior over 200–300 consecutive tool calls—far surpassing earlier versions, which degraded after just 30–50 steps. It implements a “think-before-act” approach: formulating hypotheses, selecting tools, executing actions, verifying outcomes, and iteratively refining its plan. Additionally, the model employs reasoning budgets, allowing fine-grained control over the trade-off between response accuracy and speed.
On Humanity's Last Exam (HLE)—the most challenging benchmark for analytical reasoning—K2 Thinking scores 44.9% with tool use, outperforming GPT-5 High (41.7%) and Claude Sonnet 4.5 Thinking (32.0%). On BrowseComp, a benchmark for agentic web search, it achieves 60.2%, substantially ahead of GPT-5 (54.9%) and DeepSeek-V3.2 (40.1%). In mathematical competitions, it scores 99.1% on AIME25 (with Python) and 95.1% on HMMT25, demonstrating exceptional multi-step computational reasoning. On SWE-bench Verified (real-world GitHub pull requests), it reaches 71.3%, and on SWE-bench Multilingual, it achieves 61.1%, showcasing strong programming competence with tool integration.
The model is ideally suited for autonomous research agents capable of conducting deep information analysis from web sources, performing multi-step fact verification, and synthesizing insights from dozens of documents. In software development, K2 Thinking can handle complex tasks such as refactoring, debugging, and multi-file codebase modifications using Bash, code editors, and interpreters. It demonstrates expert-level performance in solving Olympiad problems, running numerical experiments with result validation, and generating formal proofs—making it a powerful tool for scientific research and mathematical modeling. For content creation, the model excels at long-horizon writing workflows that involve research queries, claim verification, and source-cited document generation. Thanks to its stability over hundreds of tool calls, K2 Thinking is particularly well-suited for financial research requiring multi-stage analysis of financial statements and market data.
| Model Name | Context | Type | GPU | Status | Link |
|---|
There are no public endpoints for this model yet.
An endpoint is a server with pre-installed software that automatically downloads neural network weights and launches an OpenAI-compatible API to handle incoming requests.
A private endpoint is created within the user's project, belongs exclusively to the user, and is billed per server configuration in accordance with the infrastructure pricing.
We recommend creating private endpoints when you need to:
Contact our dedicated neural networks support team at support@immers.cloud or send your request to the sales department at sale@immers.cloud.