[{"_1":2,"_103":-5,"_104":-5},"loaderData",{"_3":4,"_10":11},"root",{"_5":-5,"_6":7},"user","ENV",{"_8":9},"GA_MEASUREMENT_ID","G-FZLJC5WJDF","routes/tag.$slug",{"_12":13,"_92":93,"_100":81,"_101":66,"_102":66},"posts",[14,74],{"_15":16,"_17":18,"_19":20,"_21":22,"_23":24,"_25":26,"_27":28,"_29":30,"_31":32,"_33":34,"_35":36,"_37":38,"_39":-5,"_40":41,"_42":43,"_53":54,"_67":68,"_69":70,"_71":72,"_73":36},"documentId","ahw3wecdjpaxzjvxadh5xsbb","title","Moonshot AI Kimi 'Context Caching' Feature Starts Public Beta","slug","moonshot-ai-kimi-context-caching-feature-starts-public-beta","excerpt","Moonshot AI officially launched the public beta of Kimi open platform's Context Caching feature.","content","\n

On July 1st, it was announced that Moonshot AI officially launched the public beta of Kimi open platform's Context Caching feature. The official statement mentioned that this technology can reduce up to 90% of the cost for developers using large text flagship models while maintaining API prices unchanged, and improve model response speed.

\n\n\n\n

According to the introduction, Context Caching is a data management technique that allows the system to pre-store a large amount of data or information that will be frequently requested. When users request the same information, the system can directly provide it from the cache without needing to recalculate or retrieve it from the original data source.

\n\n\n\n

Context Caching is suitable for frequent requests, repeated references to a large number of initial context scenes, which can reduce the cost of long text models and improve efficiency. The official statement indicates that costs can be reduced by up to 90%, and the delay of the first token can be reduced by 83%. The applicable business scenarios are as follows:

\n\n\n\n

Provide a QA Bot with a large amount of preset content, such as the Kimi API assistant.

\n\n\n\n

Frequent queries for a fixed set of documents, such as an information disclosure Q&A tool for listed companies.

\n\n\n\n

Periodic analysis of static code repositories or knowledge bases, such as various Copilot Agents.

\n\n\n\n

Instantly popular AI applications with huge traffic flow, such as the LLM Riddles.

\n\n\n\n

Agent-type applications with complex interaction rules.

\n\n\n\n

The Context Caching billing model is mainly divided into three parts: Cache creation cost, Cache storage cost, and Cache invocation cost.

\n\n\n\n

Call the cache creation interface, after successfully creating the cache, charge based on the actual number of tokens in the cache. Each token costs 24 yuan.

\n\n\n\n

During the cache survival time, cache storage fees are charged per minute. 10 yuan/token/minute.

\n\n\n\n

Cache call charges: Within the Cache lifetime, if a user successfully requests a created Cache through the chat interface and matches the content of the chat message with the active Cache, Cache call charges will be incurred based on the number of calls. 0.02 yuan per call.

\n\n\n\n

After the feature is launched, there will be a public beta test for 3 months, and the price during the public beta period may be adjusted at any time. During the public beta period, the Context Caching feature will be prioritized for Tier5 level users, while the release time for other user groups is to be determined.

\n\n\n\n

\n","publishedDate","2024-07-02T13:00:42.000Z","featured",false,"viewCount",0,"readingTime",3,"status","draft","language","en","originalWordPressId",165611,"featuredImage","featuredImageUrl","https://assets.pandaily.com/uploads/2024/07/kimi.jpg","categories",[44],{"_15":45,"_46":47,"_19":48,"_49":50,"_51":52},"ou7tkn9cq6eba9jr68hqiwch","name","AI","ai","color","#000000","description","","tags",[55,60,64],{"_15":56,"_46":57,"_19":58,"_49":59,"_51":52},"j45fwu61bg7hx57fo10s37oq","China","china","#6366f1",{"_15":61,"_46":62,"_19":63,"_49":59,"_51":52},"q5ktywk4z70rf8zxnq5bbabt","Moonshot AI","moonshot-ai",{"_15":65,"_46":66,"_19":66,"_49":59,"_51":52},"cipqssdhm3pbuuk4gple3qxz","kimi","createdAt","2025-06-24T22:19:44.858Z","updatedAt","2025-06-24T23:46:04.105Z","publishedAt","2025-06-24T23:46:04.136Z","locale",{"_15":75,"_17":76,"_19":77,"_21":78,"_23":79,"_25":80,"_27":28,"_29":30,"_31":81,"_33":34,"_35":36,"_37":82,"_39":-5,"_40":41,"_42":83,"_53":85,"_67":89,"_69":90,"_71":91,"_73":36},"tw7mqy2dnbtx4kp1azal2yim","Kimi Open Source Moonlight: 30 Billion / 160 Billion Parameter Hybrid Expert Model","kimi-open-source-moonlight-30-billion-160-billion-parameter-hybrid-expert-model","Kimi announced the launch of \"Moonlight\": a 30 billion / 160 billion parameter hybrid expert model (MoE) trained on Muon.","\n

Moonshot Kimi released a new technical report on \"Muon is scalable for LLM training\" yesterday and announced the launch of \"Moonlight\": a 30 billion / 160 billion parameter hybrid expert model (MoE) trained on Muon. Using 57 trillion tokens, it achieved better performance with fewer floating-point operations (FLOPs), thus improving Pareto efficiency boundaries.

\n\n\n\n

The team discovered that the Muon optimizer can be extended by adding weight decay, carefully adjusting the update magnitude of each parameter, and other techniques. It has the following highlights:

\n\n\n\n

These techniques allow Muon to be used out-of-the-box in large-scale training without needing hyperparameter tuning. Experimental results show that compared to AdamW which computes optimal training, Muon achieves approximately twice the computational efficiency.

\n\n\n\n

The model used in this paper is Moonlight-16B-A3B with a total parameter count of 15.29 billion and activation parameters of 2.24 billion. Using the Muon optimizer, it achieved the above results on training data with 5.7 trillion tokens.

\n\n\n\n

Our model not only surpasses current Pareto frontiers but also achieves better performance than previous models while significantly reducing required FLOPs for training.

\n\n\n\n

We have open-sourced a distributed version of Muon implementation optimized for memory usage and communication efficiency. Additionally, we have released pre-trained models, instruction-tuned models, and intermediate training checkpoints to support future research.

\n\n\n\n

\n","2025-02-24T14:41:12.000Z",2,168668,[84],{"_15":45,"_46":47,"_19":48,"_49":50,"_51":52},[86,87,88],{"_15":56,"_46":57,"_19":58,"_49":59,"_51":52},{"_15":61,"_46":62,"_19":63,"_49":59,"_51":52},{"_15":65,"_46":66,"_19":66,"_49":59,"_51":52},"2025-06-24T22:14:32.011Z","2025-06-24T23:45:26.426Z","2025-06-24T23:45:26.457Z","pagination",{"_94":81,"_95":96,"_97":81,"_98":99},"page","pageSize",12,"pageCount","total",14,"currentPage","tag","tagSlug","actionData","errors"]

Tag: #kimi

Moonshot AI Kimi 'Context Caching' Feature Starts Public Beta

Kimi Open Source Moonlight: 30 Billion / 160 Billion Parameter Hybrid Expert Model