News Analysis
Investors call Beijing’s Moonshot AI one of China’s “AI Tiger” companies — a firm at the forefront of China’s race toward global AI dominance. According to co-founder Yang Zhilin, his goal is to build foundational models to achieve artificial general intelligence (AGI).
This week, Moonshot AI released what it calls its “most capable” open-source thinking model: Kimi K2 Thinking.
Table of Contents
- All About Kimi K2 Thinking
- Kimi K2 Thinking Specs
- How Does Kimi Compare? K2 Thinking vs GPT-5 vs Claude Sonnet 4.5
- Kimi K2 Thinking's Agentic Capabilities
- Kimi K2 Thinking's General Capabilities
- How to Use Kimi K2 Thinking
All About Kimi K2 Thinking
According to company officials, Kimi K2 Thinking (sometimes referred to as just K2 Thinking) is built as an open-source “thinking agent that reasons step-by-step” while using tools.
K2 Thinking’s key features include:
- Deep Thinking & Tool Orchestration: Trained to weave chain-of-thought reasoning with tool calls, enabling autonomous tasks like research, coding and writing workflows.
- Native INT4 Quantization: Quantization-Aware Training (QAT) is used in the post-training stage to deliver a lossless 2× speed-up.
- Stable Long-Horizon Agency: K2 Thinking can execute 200-300 consecutive tool calls without human interference (prior models degraded after 30-50 steps).
Related Article: ChatGPT, Gemini or Grok? We Tested All 3 — Here’s What You Should Know
Kimi K2 Thinking Specs
| Architecture | Mixture-of-Experts (MoE) |
| Total Parameters | 1 Trillion |
| Activated Parameters | 32 Billion |
| Number of Layers (Dense layer included) | 61 |
| Number of Dense Layers | 1 |
| Attention Hidden Dimension | 7168 |
| MoE Hidden Dimension (per Expert) | 2048 |
| Number of Attention Heads | 64 |
| Number of Experts | 384 |
| Selected Experts per Token | 8 |
| Number of Shared Experts | 1 |
| Vocabulary Size | 160K |
| Context Length | 256K |
| Attention Mechanism | MLA |
| Activation Function | SwiGLU |
How Does Kimi Compare? K2 Thinking vs GPT-5 vs Claude Sonnet 4.5
According to Moonshot AI, K2 Thinking sets new records across benchmarks for reasoning, coding and agent capabilities. It was compared against OpenAI's GPT-5 (High) and Anthropic's Claude Sonnet 4.5 (Thinking).
K2 Learning achieved:
- 44.9% on Humanity’s Last Exam (with tools)
- 60.2% on BrowseComp
- 71.3% on SWE-Bench Verified
A Note About AI Evaluation Accuracy
Unlike traditional software with deterministic outputs, LLMs generate probabilistic results, meaning the same input can produce different outputs. Evaluations typically rely on standardized benchmarks (like MMLU or GPQA) but these tools have notable limitations.
Some benchmarks lack practical usability or fail to represent real-world scenarios adequately, and results can be misleading if not interpreted with proper context.
“Benchmarks are deeply political, performative and generative in the sense that they do not passively describe and measure how things are in the world, but actively take part in shaping it,” researchers noted. These benchmarks influence how AI models are trained, fine-tuned and applied, they added — practices with broad political, economic and cultural impacts.
Kimi K2 Thinking's Agentic Capabilities
According to Moonshot AI, Kimi K2 Thinking excels at:
Agentic Reasoning
K2 Thinking is set up with a diverse toolkit that allows it to plan, reason, execute and adapt across hundreds of steps. In one example, according to the model's creators, it solved a PhD-level math problem through 23 interwoven reasoning and tool calls.
Agentic Coding
K2 Thinking reasons while using tools, allowing it to integrate into software agents to complete complex, multi-step development workflows.
Some examples that Kimi K2 Thinking built from a single prompt, according to the company, include:
- A component-heavy website
- A short math explainer visualization
- A simulation of virus-attacking cells in bloodstream
- A vinyl record simulation
- Live coding music with Strudel.cc
Agentic Search and Browsing
K2 Thinking performs cycles of think → search → browser use → think → code. During those cycles, it generates and refines hypotheses, verifies information, reasons and constructs answers.
The AI model's intertwined reasoning allows it to turn vague, open-ended problems into clear, actionable subtasks.
Related Article: Reimagining Traditional Workflows With AI Agents
Webinar
Nov
20
How to Build a Solid Knowledge Foundation for AI Success
See how leading brands keep their AI honest, compliant and actually helpful.
Register
Webinar
Dec
2
From Manual to Magical: How AI Transforms CX Teams
Learn how to replace manual support processes with automation that actually delivers.
Register
Webinar
Dec
9
Rebrand. Migrate. Optimize. How to Do It All (Without Slowing Down)
Cresta leveled up site speed, design flexibility and marketer sanity (in record time). Find out how.
Register
Webinar
On demand
Fix the Content Bottleneck: Build a Better WebOps Strategy
Content stalled? Dev overloaded? You’re not the only one. Learn how streamlined WebOps bridges the publishing gap.
Watch Now
Webinar
On demand
Beyond Storage: Smarter Content, Bigger Impact with DAM + AI
Discover how the DAM + AI duo makes content smarter, stronger and more accessible.
Watch Now
Webinar
On demand
Agentic AI Playbook: Real-World Customer Service Use Cases You Can Deploy Now
Boost self-service by 30% and slash call volume by 63% with agentic AI.
Watch Now
Kimi K2 Thinking's General Capabilities
According to Moonshot AI, K2 Learning's other general capabilities include:
- Creative Writing: K2 Thinking has strong command of style and instruction, able to handle diverse tones and formats with natural fluency.
- Practical Writing: K2 Thinking reportedly follows prompts with high precision, often expanding on every mentioned point to ensure complete coverage.
- Personal & Emotion: The AI model responds with empathy when addressing personal or emotional questions, offering nuanced perspectives and actionable next steps.
How to Use Kimi K2 Thinking
You can use Kimi K2 Thinking now on Kimi.com under the chat mode.
This mode only uses a subset of tools and reduces the number of tool calls, meaning using K2 Thinking this way may not reproduce benchmark scores, according to company officials.
The full agentic mode will be available soon, which the model makers say will “reflect the full capabilities of K2 Thinking.”
The model is also accessible through the Kimi K2 Thinking API.
