Dark Side of the Moon releases open source thinking Agent model Kimi K2 Thinking with 1 trillion parameters. Its core breakthrough lies in the fact that it can continuously perform 200-300 tool calls without human intervention to complete complex multi-step tasks. The model adopts INT4 quantization technology to improve generation speed, and reduces computational redundancy by streamlining the architecture, with a training cost of $4.6 million. It outperforms GPT-5 in several benchmarks, including Intelligent Body Capability (τ²-Bench Telecom up to 93%), Integrated Reasoning (HLE up to 44.9%), and Programming Practice (SWE-Bench Verified up to 71.3%). The model is completely open source and commercially free under a modified MIT license.
- This summary is generated by AI analyzing the content of the article for informational purposes only.
"The ability to call up the tool 300 times in a row without human intervention is true thinking power."
Last night, there was an earthquake in the field of domestic large models - Dark Side of the Moon was officially released Kimi K2 ThinkingA new product with a 1 trillion parameters The open source Think Agent model.
Less than 2 hours after going live, the servers were packed; Hugging Face co-founder Thomas Wolf was thrilled, saying, "We are witnessing another DeepSeek moment."; Sebastian Raschka, a renowned AI scholar, analyzed that "More experts, less HEAD, more thinking"; Reddit Hot Topic comment:"The closest the open-source model is to the closed-source frontier.“
What is it about this model that has the global AI community buzzing? Today, we're going to find out.
Kimi K2 Thinking is not an ordinary upgrade, but a completely re-engineered Thinking Agent with only one core breakthrough:
Performs 200-300 consecutive tool calls for complex multi-step tasks without manual intervention.
Unlike traditional big models that can only be answered passively, K2 Thinking will actively think, question, validate, and adjust, thinking and acting like a human being.
norm
K2 Thinking
traditional model
Tool Call
200-300 consecutive calls
Usually 1-3 times
thought process
Explicitly demonstrate the chain of reasoning
Hide the reasoning process
interactive mode
Active Search + Think + Execute
passive response
Type of mission
Multi-step complex tasks
single simple task
⚡ Technical underpinnings: the secret behind the 1T parameter
1️⃣ Quantitative Breakthrough: INT4 is not a compromise, it's a strategy
Compared with the FP8 adopted by competitors, K2 Thinking chose to quantize INT4, which is not only a technological breakthrough, but also a strategic consideration:
Double the speed: Increase in generation speed by about 2 times
hardware compatibility: More friendly to domestic accelerated computing chips
No loss of performance: Maintaining performance without degradation through Quantitative Awareness Training (QAT)
"Two Macs with M3 Ultra chips can run INT4 Compressed smoothly with little to no performance loss." --Awni Hannun, Apple Bully, Test Share
2️⃣ Architectural revamp: more experts, less head
Compared to DeepSeek R1, K2 Thinking utilizes a more streamlined architecture:
More Experts: Enhancing the breadth of model knowledge
Less head: Reduce computational redundancy
alternate thinking: Cycle between "think" and "do" to improve reasoning coherence
3️⃣ Training Cost: $4.6 Million Precision Investment
According to CNBC, K2 Thinking costs $4.6 million to train. Compared to the tens of millions of dollars of model training, Dark Side of the Moon maximizes performance with a precise and efficient training strategy. Every penny is spent on the knife edge.
📊 Performance in action: surpassing GPT-5 SOTA scores
The K2 Thinking has demonstrated amazing strength in a number of authoritative benchmark tests:
🔍 Intelligent body capabilities: truly autonomous thinking
𝜏²-Bench Telecom:: 93% accuracy over GPT-5 (89%) and Claude Sonnet 4.5 (91%)
SEAL-0: Complex Information Gathering Reasoning Test, Refreshing SOTA
BrowseComp:: 60.21 TP3T score, human average only 29.21 TP3T
Generate a moving map: blue contour lines, yellow paths, red gradient arrows
With textual explanations, at a glance
Parameters can be adjusted interactively by the user
✅ Case 3: Viral Spread Simulation
mandates: Make a virus simulation program with adjustable immune parameters
Performance of K2 Thinking::
23 tool calls
Generate a fully interactive program
Red and blue particles chasing, colliding and devouring each other
Sliders to adjust viral replication rate, number of immune cells
Real-time parameter feedback and statistics
✅ Case 4: Data Analysis and Visualization
mandates:: "Analyze the CSV file I sent you and generate charts to support the analysis"
Performance of K2 Thinking::
First planning step: load data → screening → analysis → mapping
14 Python calls
Generate interactive web pages with statistical analysis, visualization charts, detailed explanations
Error self-healing without human intervention
🚀 Free and open source: the AI revolution available for all
Most excitingly, K2 ThinkingCompletely open sourceAdoptionModified MIT License::
✅ Commercial Free: Can be used directly in commercial products
✅ model weight: Full Open on Hugging Face
✅ API interface: Services provided by Kimi Open Platform
✅ personal useAvailable instantly on kimi.com and mobile apps!
The only restriction: Kimi K2 needs to be prominently labeled in the UI when there are more than 100 million monthly active users or more than $20 million in monthly revenue.
"This is not a victory for one company, but a collective leap in China's AI ecosystem." -- CTO of a head AI company
🌟 Write in the end
The emergence of Kimi K2 Thinking breaks our inherent perception of AI. It is no longer a passive "chatbot", but a "digital colleague" that can actively think, solve problems and continuously evolve.
As open-source models begin to overtake closed-source models, and as Chinese technology begins to lead global AI innovation, we have to admit:China is at the forefront of the AGI journey.
"It's not replacing humans, it's liberating them. Let AI handle the tedious calculations and execution, and humans focus on creation and decision-making." --Dark Side of the Moon Engineer Team
Ali Tongyi Thousand Questions team released a new generation of open source large model Qwen3, topped the global open source model list. The series contains models, the flagship model performance exceeds a number of top models, deployed significantly...
MiniMax releases M2, a new generation of open source large model with top 5 performance worldwide, priced at only 8% of Claude 4.5.The model has 230B total parameters, activation...
November 11, 2025 - China's open-source model Kimi K2 has sparked industry interest by outperforming the GPT-5 on a number of tasks at a quarter of the latter's cost. Meanwhile...
Comments are closed.