What is Kimi Latest
Kimi Latest is a real-time updated AI model launched by Moonshot AI's Kimi, providing users with a synchronized benchmarkingKimi intelligent assistantmodel experience. It supports a 128k context length and can automatically select 8k, 32k, or 128k models for billing based on input, while also possessing image understanding capabilities to handle visual tasks. Kimi Latest supports automatic context caching, and the cost of Tokens that hit the cache is extremely low, only 1/M Tokens.
Main features of Kimi Latest
- Latest model capabilities: Always benchmarks against the latest model currently used by the Kimi intelligent assistant, including experimental features that are not yet stable.
- Ultra-long context support: Supports a 128k context length and can automatically select 8k, 32k, or 128k models for billing based on input.
- Image understanding capability: As a vision model, it supports image understanding functionality.
- Automatic context caching: Supports automatic context caching, and the cost of Tokens that hit the cache is only 1/M Tokens.
- Advanced feature integration: Supports ToolCalls, JSON Mode, Partial Mode, web search functionality, and more.
- Wide range of applicable scenarios: Suitable for large model chat applications (such as ChatWise, ChatBox, etc.), providing a chat experience similar to the Kimi intelligent assistant.
Technical principles of Kimi Latest
- Long Chain-of-Thought technology (Long-CoT): The Kimi Latest model inherits the Long Chain-of-Thought technology of Kimi K1.5. By extending the length of the reasoning chain, the model can perform deeper thinking and complex reasoning. Compared with traditional short reasoning chains, the long chain of thought can significantly improve the model's performance in mathematics, programming, and multimodal reasoning tasks.
- Long Context Scaling: Kimi Latest supports a context length of up to 128k and optimizes computational efficiency through the Partial Rollout strategy. The model has greater advantages when processing long texts and complex dialogues.
- Improved Policy Optimization: The model adopts a reinforcement learning formula based on long chain-of-thought, combined with Online Mirror Descent for policy optimization.
- Multimodalities joint training (Multimodalities): Kimi Latest supports multimodal information processing, including joint training on text and visual data. This enables the model to process text and visual information simultaneously and achieve cross-modal reasoning.
- Automatic context caching: Kimi Latest introduces automatic context caching technology, and the cost of Tokens that hit the cache is only 1/M Tokens.
- Concise framework design: The model abandons complex technologies such as Monte Carlo tree search and value functions, and achieves powerful reasoning capabilities by extending context length and optimizing strategies.
- Long2Short technology: Kimi Latest adopts Long2Short technology, transferring the reasoning capability of the long-chain thinking model to the short-chain thinking model, further optimizing performance.
Project address of Kimi Latest
- Project official website：Kimi intelligent assistant
Model pricing of Kimi Latest
- kimi-latest-8k: 1M tokens; input price (cache hit) ¥1.00; input price (cache miss) ¥12.00; output price ¥12.00.
- kimi-latest-32k: 1M tokens; input price (cache hit) ¥1.00; input price (cache miss) ¥24.00; output price ¥24.00.
- kimi-latest-128k: 1M tokens; input price (cache hit) ¥1.00; input price (cache miss) ¥60.00; output price ¥60.00.
Why launch Kimi Latest?
- Meet differentiated needs: The user needs of the Kimi intelligent assistant and the open platform are gradually diverging. The intelligent assistant needs replies with greater emotional value, while open platform users place more emphasis on the model's structured processing and precise output of text. This difference leads to problems such as prompt failure that users may encounter during use.
- Maintain model stability: The rapid iteration characteristics of the Kimi intelligent assistant can continuously launch new features, but may also introduce some experimental features that are not yet stable. Although these features perform excellently in some scenarios, they may have problems in other scenarios. Therefore, differences exist between the models on the open platform and the models used by the intelligent assistant, raising users' doubts about inconsistent replies across different platforms.
- Synchronize the latest capabilities: To solve the above problems, the Kimi Latest model is launched. It will always align with the latest model currently used by the Kimi intelligent assistant and upgrade synchronously with updates to the intelligent assistant. In this way, open platform users can both experience the excellent results brought by the latest model and maintain the stability of the original model.
- Improve user experience: By launching the Kimi Latest model, Moonshot AI provides developers and users with a more flexible, efficient, and stable AI model solution to meet their diverse needs in different scenarios.
Special notes on Kimi Latest
- If you are using large model chat applications such as ChatWise, ChatBox, LobeChat, OpenCat, and want to use Kimi as your AI chat partner, the Kimi Latest model is the best choice.
- To build an AI intelligent assistant or customer service for a product and pursue a chat experience similar to the Kimi intelligent assistant, choose the Kimi Latest model.
- To use a large model for intent recognition or structured data extraction, please choose the moonshot-v1 series models.
- If the moonshot-v1 series models perform stably, you can continue to use them; if you switch to the Kimi Latest model, you need to revise the prompts again.
Application scenarios of Kimi Latest
- Language learning: Help users practice language skills and provide translation, grammar correction, and writing guidance.
- Teaching assistance: Help teachers design course outlines, create teaching materials, and evaluate student assignments.
- Market research: Analyze market trends and competitor information, and provide market research reports.
- Content creation: Generate copy, advertising slogans, news reports, etc., and improve content creation efficiency.
- Image recognition and analysis: Recognize objects, scenes, and text in images, and provide image descriptions and analysis reports.