Run Kimi K2, the latest open-source SOTA model, on W&B Inference
Get started with Kimi K2 Instruct on our free tier
Created on July 15|Last edited on July 16
Comment
W&B Inference powered by CoreWeave now provides API and playground access to the Kimi K2 Instruct model, enabling developers to build agentic AI applications without deploying on their own or managing multiple model provider keys.
Kimi K2 is Moonshot AI’s open-source mixture-of-experts model with 1 trillion parameters optimized for agentic tasks, delivering state-of-the-art frontier knowledge, math, and coding capabilities. You can now instantly access Kimi K2, hosted on CoreWeave Cloud Platform, with automatic tracing for seamless observability in W&B Weave. Evaluate, monitor, and iterate your agentic AI applications quickly, leveraging integrated access to W&B Weave tracing when using Kimi K2 via W&B Inference.
💡
What’s Kimi K2 Instruct?
Kimi K2 is Moonshot AI’s latest open-source “agentic intelligence” model featuring a mixture-of-experts architecture with 32 billion activated parameters and 1 trillion total parameters. Unlike typical models, Kimi K2 doesn’t just answer queries. It acts using tools, running commands, editing files, and orchestrating complex workflows autonomously. It’s specifically optimized for agentic tasks, enabling seamless integration into practical applications without extensive manual workflow setups.
Why does it matter?
What sets Kimi K2 apart is its outstanding performance in coding, math, STEM, and real-world agentic interactions. Benchmarks show it consistently matches or beats top proprietary and open-source models, especially shining in tough tasks like competitive coding, multilingual software development, and complex tool-driven workflows.
Leveraging new optimization techniques such as the MuonClip optimizer, it achieves remarkable training stability and efficiency, enabling large-scale, stable training of powerful agentic intelligence.
Figure 2: Pre-training loss curve (source: Moonshot AI)
The community excitement around Kimi K2 comes from how it democratizes access to powerful, actionable AI, closing the gap between theory and practical agentic use. Its powerful combination of superior task performance, open availability, and ease-of-use marks a significant advancement in building accessible, real-world intelligent applications.
Server-less inference on W&B Inference powered by CoreWeave
Skip the hassle of signing up for one more model hosting provider or deploying the model yourself. Your Weights & Biases account gives you instant access to Kimi K2 Instruct and other top open-source foundation models, fully hosted on powerful CoreWeave infrastructure. Simply sign in to Weights & Biases, select Kimi K2 Instruct from the menu, and start running inference for free in seconds.
You can also try Kimi K2 instantly in the W&B Weave Playground—no model endpoints, no access keys, and zero configuration needed.
If you want to run it from your code, navigate to the model card and copy and paste the starter code we’ve provided.
Integrated observability
Agentic AI applications need observability tools, but model hosting providers may not offer them, forcing developers to juggle disconnected platforms for hosting and observability. W&B Inference runs on CoreWeave Cloud Platform with observability built-in through W&B Weave to evaluate, monitor, and iterate on AI applications and agents—no additional instrumentation, fragmented workflows, or complexity.
While W&B Weave offers a built-in integration for observability into LLM calls made through W&B Inference, it’s entirely optional. W&B Inference works independently and can be used on its own if your focus is solely on getting fast, scalable inference. The two tools are complementary but not coupled.
Evaluating Kimi K2 for your use case
Best of all, you can quickly compare Kimi K2 against other open-source models for your specific use case without juggling multiple provider accounts or API keys. Just bring your favorite prompts and see how Kimi K2 stacks up side-by-side with other open-source models.
You can easily decide if switching to Kimi K2 gives you comparable quality and performance at a fraction of the cost of proprietary alternatives. Test interactively in the playground or run offline evaluations on your dataset using the W&B Weave Evaluations API. You can also use Kimi K2 on W&B Inference for your LLM-as-judge scorer for evaluations, monitors, and guardrails.
Getting started
Try Kimi K2 Instruct right now through the W&B Weave Playground. Every Weights & Biases plan includes a free tier of W&B Inference, so you can dive straight in without additional upfront costs. To learn more, see the W&B Inference documentation and the W&B Inference pricing page.
Add a comment
Iterate on AI agents and models faster. Try Weights & Biases today.