Kimi Models: Performance and Selection Guide
Master the full spectrum of Moonshot AI's powerful 1 trillion parameter models for any workflow.
Explore Models
Kimi Models Overview
Moonshot AI's Kimi platform offers an impressive lineup of AI models built on a 1 trillion parameter Mixture-of-Experts architecture. The flagship Kimi K2.5, released in January 2026, delivers native multimodal capabilities, Agent Swarm coordination, and benchmark performance that rivals GPT-5.2 and Claude Opus 4.5. Whether you need fast responses for simple queries, deep reasoning for complex problems, or autonomous agent workflows, the Kimi model family covers every use case.
The Kimi ecosystem has evolved rapidly since K1.5 launched in January 2025. Each subsequent release has expanded capabilities from text-only reasoning to full multimodal understanding with video, images, and documents. All K2-series models share the same 1T MoE foundation but differ in training data, feature sets, and operational modes. The entire K2.5 model is open-source under a Modified MIT License, available on Hugging Face for self-hosted deployment.
| Model | Release Date | Parameters | Context Window | Key Features |
|---|---|---|---|---|
| Kimi K2.5 | January 2026 | 1T MoE (32B active) | 256K tokens | Native multimodal, Agent Swarm, open-source |
| Kimi K2-Instruct-0905 | September 2025 | 1T MoE (32B active) | 256K tokens | Improved coding, extended context |
| Kimi K2 | July 2025 | 1T MoE (32B active) | 128K tokens | First 1T MoE, open-source base |
| Kimi Linear | October 2025 | 48B MoE (3B active) | 128K tokens | Lightweight, efficient inference |
| Kimi-VL | April 2025 | 16B MoE (3B active) | 128K tokens | Vision-language, compact multimodal |
| Kimi K1.5 | January 2025 | Undisclosed | 128K tokens | Reasoning parity with OpenAI o1 |
Kimi K2.5 Flagship Model
Kimi K2.5 represents the most capable model in the lineup, trained on approximately 15 trillion mixed visual and text tokens through continual pre-training atop the K2 base. The architecture uses 384 experts with 8 activated per token, Multi-Latent Attention (MLA), and SwiGLU activation. The native multimodal design integrates MoonViT-3D, a 400M parameter vision encoder using NaViT packing for variable-resolution image input.
Four Operational Modes
K2.5 operates in four distinct modes, each optimized for different workflows. K2.5 Instant delivers fast, non-thinking responses for straightforward queries. K2.5 Thinking activates chain-of-thought reasoning for complex problems. K2.5 Agent enables single-agent tool use for autonomous task completion. K2.5 Agent Swarm coordinates up to 100 specialized sub-agents working in parallel, cutting execution time by 4.5x.
| Mode | Use Case | Speed | Reasoning Depth |
|---|---|---|---|
| K2.5 Instant | Quick answers, simple tasks | Fastest | Standard |
| K2.5 Thinking | Math, logic, complex analysis | Moderate | Deep chain-of-thought |
| K2.5 Agent | Tool use, code execution, browsing | Task-dependent | Agentic reasoning |
| K2.5 Agent Swarm | Complex research, multi-step workflows | 4.5x faster than single agent | Distributed multi-agent |
Benchmark Performance
K2.5 achieves 96.1% on AIME 2025 (GPT-5.2: 100%), 98.0% on MATH-500, and 87.6% on GPQA-Diamond. In coding, it scores 83.1% on LiveCodeBench v6, significantly outperforming Claude Opus 4.5's 64.0%. The Agent Swarm mode achieved 50.2% on Humanity's Last Exam with tools, surpassing GPT-5.2's 45.5% at 76% lower cost. Vision capabilities include 92.3% on OCRBench and 86.6% on VideoMMMU.
Kimi K2 Base Model
Released in July 2025, K2 was Moonshot AI's first 1 trillion parameter MoE model and the foundation for all subsequent K2-series releases. Open-sourced under MIT License, it established the 384-expert architecture with 32B activated parameters that K2.5 inherits. The original release supported 128K tokens of context, later extended to 256K with the September 2025 Instruct update.
K2-Instruct-0905, released in September 2025, brought significant coding improvements and the expanded 256K context window. This update scored 94.5% on HumanEval, demonstrating strong code generation capabilities. The Instruct variant remains available as a text-only alternative for users who do not need multimodal features.
Lightweight Models
Kimi Linear
Launched in October 2025, Kimi Linear uses a compact 48B MoE architecture with only 3B activated parameters per token. Designed for edge deployment and resource-constrained environments, it delivers surprisingly capable performance relative to its size. The model supports 128K token context and runs efficiently on consumer-grade hardware, making it suitable for local deployment, mobile applications, and high-throughput scenarios where latency matters more than peak capability.
Kimi-VL
Released in April 2025, Kimi-VL is a 16B parameter MoE vision-language model with 3B activated parameters. It was Moonshot AI's first open-source multimodal model, designed for tasks combining image understanding with text generation. While superseded by K2.5's native multimodal capabilities for demanding workloads, Kimi-VL remains valuable for lightweight vision tasks where the full 1T model would be excessive.
Kimi K1.5 Reasoning Model
K1.5, released in January 2025, marked Moonshot AI's entry into advanced reasoning models. It claimed performance parity with OpenAI o1 on math and coding benchmarks, introducing reinforcement learning-based reasoning capabilities to the Kimi platform. While the exact parameter count was never disclosed, K1.5 demonstrated that Moonshot AI could compete at the frontier of AI reasoning.
K1.5 focused exclusively on text-based reasoning without multimodal capabilities. Its release established Moonshot AI as a serious competitor in the reasoning model space and laid the groundwork for the more capable K2 series that followed. Users still on K1.5 should upgrade to K2.5, which surpasses K1.5 in every benchmark while adding multimodal and agentic capabilities.
Choosing the Right Kimi Model
Model selection depends on your specific requirements for capability, cost, and deployment flexibility. The following guide helps match use cases to the optimal model choice.
- General-purpose AI with maximum capability: Use K2.5 via API or kimi.com. The Instant mode handles simple tasks efficiently, while Thinking mode tackles complex reasoning.
- Autonomous workflows and research: Use K2.5 Agent or Agent Swarm mode. The swarm system excels at multi-step tasks requiring parallel information gathering.
- Self-hosted deployment with full features: Download K2.5 from Hugging Face (moonshotai/Kimi-K2.5) and deploy via vLLM, SGLang, or Docker.
- Lightweight or edge deployment: Use Kimi Linear (48B MoE, 3B active) for resource-constrained environments requiring capable AI on modest hardware.
- Simple vision tasks on a budget: Use Kimi-VL (16B MoE, 3B active) when full K2.5 multimodal capabilities exceed your needs.
- Text-only reasoning at lower cost: Use K2-Instruct-0905 when you do not need vision capabilities but want the 1T MoE reasoning power.
Frequently Asked Questions
Which Kimi model is the best?
Kimi K2.5 is currently the most capable model, excelling in reasoning, coding, vision, and agentic tasks with top-tier benchmark scores.
Are Kimi models free to use?
Yes, they are free to use on kimi.com and mobile apps. API access and open-source deployment are also available with specific pricing and licenses.
Can I run Kimi models locally?
Yes, models like K2.5 and Kimi Linear are available on Hugging Face for local deployment, though K2.5 requires significant GPU resources.
What is the difference between K2 and K2.5?
K2.5 introduces native multimodal support, Agent Swarm mode, and higher performance benchmarks compared to the text-only K2.
What is Agent Swarm mode?
It is a feature of K2.5 that coordinates up to 100 specialized sub-agents to complete complex research and workflows 4.5x faster.
What is Kimi Linear designed for?
Kimi Linear is a lightweight 48B MoE model optimized for edge deployment and high-throughput scenarios where low latency is critical.
Has Kimi K1.5 been surpassed?
Yes, K2.5 benchmarks higher in every category and adds multimodal and agentic capabilities that K1.5 lacked.
What license do Kimi models use?
The K2-series and K2.5 weights are released under a Modified MIT License, allowing for both exploration and commercial use.