What is Kimi K2 Thinking
Kimi K2 Thinking is Moonshot AI’s latest and most powerful open‑source “thinking” model. It is designed as a smart agent that can think step by step and invoke different tools as it works.
This model is available in Microsoft Foundry as a Direct from Azure model, which means Microsoft hosts and manages it directly on Azure.
What does “Direct from Azure” mean
“Direct from Azure” models are a special set of models that Microsoft sells and manages itself.
They are secured and operated by Microsoft, so you get enterprise‑grade security, support, and reliability.
Billing, governance, and usage are all handled in one place inside Microsoft Foundry, which makes operations simpler.
You can easily try, deploy, or switch between models on the same platform as new models are added.
You can control costs with pay‑as‑you‑go or reserved capacity, depending on your needs.
Key capabilities of Kimi K2 Thinking
Kimi K2 Thinking focuses on deep reasoning and long, stable workflows.
Deep thinking and tool orchestration: It is trained to mix step‑by‑step reasoning (chain‑of‑thought) with function calls, so it can run long workflows for research, coding, or writing without losing track of the goal.
Native INT4 quantization: It uses a special low‑precision format (INT4) with Quantization‑Aware Training to speed up inference about 2x, without losing quality, and to save GPU memory.
Stable long-horizon agency: It can keep following the same objective across 200–300 tool calls in a row, while many older models start to break down after around 30–50 steps.
It also sets new state‑of‑the‑art results on benchmarks like Humanity’s Last Exam (HLE) and BrowseComp, which test difficult reasoning and multi‑step tool use.
Architecture and technical specs
Under the hood, Kimi K2 Thinking is a very large Mixture‑of‑Experts (MoE) model.
Total parameters: about 1 trillion, but only 32 billion are “active” for each token thanks to MoE routing.
It has 61 layers in total, with 1 dense layer and many expert layers.
It uses 64 attention heads, with hidden dimensions tuned for large‑scale reasoning.
There are 384 experts, and 8 experts are selected per token, plus 1 shared expert.
Context length is 256K tokens, so it can handle very long conversations or documents in one go.
Activation function is SwiGLU, and attention uses an MLA‑style mechanism.
Some details like training cut‑off date, training time, supported languages, and input/output formats are not provided in the catalog yet.
Join the conversation! Your thoughts help the community grow.