Moonshot AI's Kimi K2 Released and Open-Sourced, Perplexity CEO: "Stunning! Post-Training About to Begin" | BlueRun Ventures Family Headlines

On July 11, Moonshot AI officially released the Kimi K2 model, simultaneously open-sourcing it.

The K2 model has received widespread acclaim across the industry. Just now, the Perplexity CEO posted on X: $2*$2$2$2$2$2$2$2**$2*

BlueRun Ventures was an early investor in Moonshot AI and has continued to increase its stake in subsequent rounds. We congratulate Moonshot AI on reaching this new milestone.

---

Kimi K2 is a Mixture-of-Experts (MoE) foundation model with stronger coding capabilities and greater proficiency at general agent tasks, featuring 1 trillion total parameters and 32 billion active parameters.

On benchmark tests including SWE Bench Verified, Tau2, and AceBench, Kimi K2 achieved state-of-the-art results among open-source models, demonstrating leading capabilities in coding, agent tasks, and mathematical reasoning. During pre-training, Kimi K2 used the MuonClip optimizer to enable stable and efficient training of a trillion-parameter model. Amid the bottleneck of high-quality human data, this approach effectively improved token utilization efficiency and uncovered new scaling opportunities.

Other key technologies include large-scale agentic tool use data synthesis and general reinforcement learning with self-evaluation mechanisms. For more details, please refer to our technical blog.

Starting today, visit kimi.com or download the Kimi App to experience the all-new Kimi K2 model. API services are also now live, offering Chat API interfaces compatible with OpenAI and Anthropic formats. You can easily switch your existing LLM tools to Kimi K2 and experience its powerful agent and coding capabilities.

Kimi K2 provides a solid foundation for building general agent capabilities, but truly general agents will require more advanced abilities, such as reasoning and visual understanding. We plan to add these capabilities to Kimi K2 in the future.

We hope that by fully open-sourcing a more capable model, we can further accelerate the overall progress of AGI research and real-world application.

Kimi K2 achieved excellent results in benchmark performance tests across three capability dimensions: agentic coding, tool use, and math & reasoning.

Beyond benchmark tests, Kimi K2 also demonstrates stronger capability generalization and practical utility across multiple real-world scenarios:

Enhanced Coding Capabilities

In front-end development tasks, Kimi K2 excels at generating code with both design sensibility and visual impact, supporting particle systems, data visualizations, and 3D scenes, with strong graphics capabilities and interactivity.

Below is a 3D mountain canyon landscape generated by Kimi K2, featuring day-night cycles:

$2

Here is a particle-effect galaxy generated by Kimi K2:

$2

** This is a futures trading system generated by Kimi K2 in one shot. Without specific instructions, Kimi automatically selected TradingView and built a complete futures trading interface: $2

Enhanced Agent Tool Use Capabilities Kimi K2 now features stable complex instruction parsing, automatically breaking down requirements into a series of well-formatted, directly executable ToolCall structures.

You can seamlessly integrate it with agent/coding frameworks such as owl, Cline, and RooCode to complete complex tasks or automated coding.

Agent capabilities are already available via API, with more tool features coming soon to the Kimi platform. First, take a look at this real demo from our internal testing environment to experience the power of a model with strong agentic capabilities:

For example, feed Kimi K2 130,000 rows of raw data, and it can analyze how remote work ratios affect salaries, identify statistically significant differences, automatically generate statistical charts and regression model interpretations, and produce professional visualizations — violin plots, box plots, scatter plots — in a unified color scheme, compiling everything into a report.

Or, if you're a Coldplay fan, Kimi K2 can help you plan this year's concert tour, complete with flight and hotel bookings and travel itineraries for each city, generate a calendar, and then summarize the full itinerary in HTML and email it to you.

Enhanced Stylized Writing Capabilities

In rewriting tasks, Kimi K2 can precisely control output style. Whether rewriting scientific text in a middle-schooler's voice or mimicking Apple advertising copy, it preserves both the original meaning and the target expression style, demonstrating strong context retention and expressive transfer abilities.

In creative fiction tasks, Kimi K2 generates text that pays more attention to detail and emotion, rather than speaking in vague abstractions.

When we gave Kimi K2 a science fiction writing prompt that once sparked widespread discussion — "What if the real world is actually an AI model?" —

Kimi K2 generated a richly plotted science fiction story full of detailed descriptions, with passages that prove genuinely moving:

$2

$2

Below is the full text of the work Kimi K2 generated based on this premise:

Additionally, Kimi K2 shows improvement in tasks involving general knowledge reasoning, mathematics, and planning.

We have simultaneously open-sourced two model versions from the Kimi K2 series:

• Kimi-K2-Base: The base pre-trained model without instruction fine-tuning, suitable for research and custom scenarios;
• Kimi-K2-Instruct: The general instruction-tuned version (non-reasoning model), delivering excellent performance on most Q&A and agent tasks.

Model weights and fp8 weight files have been open-sourced on Hugging Face 👇

https://huggingface.co/moonshotai/Kimi-K2-Instruct

Additionally, inference engines including vLLM, SGLang, and ktransformers have added support, allowing you to deploy on your own servers for an experience identical to the Kimi Open Platform API.

Kimi K2 uses the MuonClip optimizer to robustly support trillion-parameter model training, significantly improving token utilization efficiency. Combined with large-scale agentic data synthesis and general reinforcement learning, the model continues to advance in general intelligence capabilities.

• MuonClip Optimizer: Kimi K2 abandoned the traditional Adam optimizer, innovatively adopting the Muon optimizer. To mitigate the issue of attention logits growing too large during large-scale training, we proposed MuonClip and scaled it to trillion-parameter dimensions, improving training stability and token efficiency. Kimi K2 completed stable training on 15.5 trillion tokens with no loss spikes throughout.
• Large-Scale Agentic Tool Use Data Synthesis: We built a synthetic pipeline capable of generating multi-turn tool use scenarios at scale, covering hundreds of domains and thousands of tools. High-quality samples were evaluated and filtered by LLMs for training use.
• General Reinforcement Learning: Kimi K2 applies reinforcement learning not only to verifiable tasks (code, math) but also solves the reward sparsity problem for unverifiable tasks by introducing self-judging mechanisms. By continuously optimizing the critic on verifiable tasks, generalization performance on broader tasks is improved.

Kimi K2's API service is now fully live, supporting context lengths up to 128K with greater versatility and tool use capabilities. Pricing is as follows:

• 4 RMB per million input tokens
• 16 RMB per million output tokens

We are compatible with both OpenAI and Anthropic API formats, and integrate well with various frameworks. Additionally, the newly upgraded ToolCall capabilities strictly guarantee format correctness, making them suitable for complex agent tasks.

See more 👉 Kimi Open Platform

Visit kimi.com or download the Kimi App to start a conversation with the Kimi K2 model right away.

(Tool use capabilities are already built into the model; related features are currently in internal testing and will be released soon. Stay tuned!)

If you'd like to join us in exploring Moonshot AI, in exploring the optimal solution for converting energy into intelligence, we welcome your application 👉 kimi

Kimi Researcher Team: Agent Is Not a Frankenstein, It Is the Model | BlueRun Family Headlines BlueRun Family Headlines | Kimi Partners with Tsinghua University and Other Institutions to Open-Source Mooncake, a Large Model Inference Architecture BlueRun in Conversation with Moonshot AI and Muyan Zhiyu: Standing at the Crossroads of the AI Industry | The Road to AGI