Kimi

Overview

Kimi is Moonshot AI's assistant, first known to Chinese users for long-document handling. The lineup now splits three ways: the consumer assistant, the Kimi Code and K2.7 Code open models for developers, and Kimi Work for enterprises.

On model generations, K3, released on July 16, 2026, was the turning point. It carries 2.8 trillion total parameters in a sparse MoE layout that activates 16 of 896 experts, the in-house KDA hybrid linear attention, and a 1 million token context. Weights went out on July 27 as promised, making it the largest open-weight model at the time. K3.1 Pro followed on October 23 as a post-trained version of K3 with the same total parameter count, 104 billion active parameters, and a score of 55 on the Artificial Analysis Intelligence Index, fourth worldwide, up from 44 for K3.

Commercially, annualized recurring revenue passed 300 million dollars in mid June, after crossing 100 million in March and 200 million in May. Moonshot closed a funding round at a 20 billion dollar valuation on June 30, and a new round lifted the pre-money valuation to 31.5 billion dollars. Basic chat and file analysis on the web and in the app remain free.

Key Features

• 1 million token context: A single conversation can take in a full novel, an entire earnings report or a set of contracts, so cross-chapter comparison does not require splitting uploads. Long documents are where Kimi first earned its reputation, and K3 carries that ability into the new architecture.
• K3.1 Pro (new): A post-trained version of K3 released on October 23, 2026. 2.8 trillion total parameters, 104 billion active, 55 on the Artificial Analysis Intelligence Index for fourth worldwide, 65.7 percent on Terminal-Bench 4.0, and API pricing at 2 dollars input and 10 dollars output per million tokens.
• Open K3 weights: 2.8 trillion parameters, 16 of 896 experts active, KDA hybrid linear attention and native vision understanding. Weights shipped on July 27, 2026 as announced, so the model can be downloaded and run on your own cluster.
• Kimi Code and K2.7 Code: Two coding lines. K2.6 targets 13 hour continuous coding runs and orchestration across 300 agents, while K2.7 Code is tuned specifically for programming work and pairs with the Kimi Code client for edits inside a repository.
• Kimi Work: The enterprise line, connecting document processing, internal search and a knowledge base. Priced separately by seat and usage, with its own account system apart from the consumer assistant.
• Web search and file parsing: Paste a link and it pulls the article text. PDF, Word and Excel files are analyzed page by page with cross references, and papers can be broken down by section to extract methods and conclusions. Search results carry source links for checking.

Use Cases

• Read through a full novel, a long report or dozens of contracts and pull cross-chapter terms, character relationships or figures into one table.
• Compare earnings reports and research notes across several companies in a single conversation instead of switching files repeatedly.
• Work through papers and patents section by section, extracting methods, datasets and conclusions, then ask the model to point out where the argument stays vague.
• Deploy K3 weights on your own cluster for internal document Q&A or code completion without data leaving the building.
• Use Kimi Code for repository-wide refactors and test coverage, where agent orchestration can split a long chain of changes across parallel instances.

Pros

• A 1 million token context takes whole books in one pass, so long documents need no pre-splitting.
• K3 weights were released as promised, and a 2.8 trillion parameter model can be deployed locally.
• K3.1 Pro API pricing sits at 2 dollars input and 10 dollars output, roughly a third below K3.
• Basic chat and file analysis stay free on the web and in the mobile app.
• Chat, coding and enterprise work each have their own entry point, so you can adopt only the piece you need.

Pricing

Basic chat and file analysis are free on the web and in the mobile app. API pricing runs per token: K3.1 Pro at 2 dollars input and 10 dollars output per million tokens, Highspeed at 4 dollars input and 20 dollars output, K3 at 3 dollars input (0.30 dollars on cache hits) and 15 dollars output. K3 weights are open, so self-hosting carries only compute cost. Kimi Work is quoted separately for enterprises and billed apart from consumer accounts. Check the official site for current quotas, concurrency and rate limits.

Summary

Kimi built its reputation on long documents, and after K3 the focus moved to the models themselves. Open weights at 2.8 trillion parameters, a 1 million token context and K3.1 Pro at fourth on the intelligence index add up to a product that works as a daily tool and as a model base. Reading papers, pulling apart earnings reports and drafting long pieces still feel natural, and for your own applications the weights and the API are both within reach. Anyone who still treats it as a free long-text tool should look at which features in the app have already moved to K3.1 Pro.

Version History

• Introducing Kimi K3 on Amazon Bedrock (2026-09-18): Kimi K3 from Moonshot AI is now available on Amazon Bedrock, giving you a powerful new open-weight option for coding and knowledge work. It offers native vision, a 1 million token context window, and
• K3 open-source commercial revenue-sharing rules (2026-08-07): K3 sets a 30% commercial revenue-sharing threshold, Qwen3.8-Max beats K3 on some benchmarks; developers worldwide continue deploying the K3 ecosystem
• WorkBuddy deeply integrates with Tencent Docs, building third-generation office collaboration for the Agent era (2026-07-30): WorkBuddy's new version deeply integrates with Tencent Docs, supporting direct editing of generated files in the sidebar and embedding Tencent Docs into workflows to enable collaboration between humans, AI, colleagues, and their respective Agents. Users can choose models such as DeepSeek V4 Pro and Kimi K3 to process documents, and upload files to the cloud for sharing and circulation. The author believes that this collaboration model of "humans, Agents, and shared state" marks the arrival of the third generation of office products.
• Running the 2.8T-parameter Kimi K3 on an M1 Max: the Deltafin project achieves 0.0687 token/s inference (2026-07-29): The Deltafin project successfully ran the 2.8T-parameter MoE model Kimi K3 on a 64 GB M1 Max, with a current median inference speed of 0.0687 tokens/s (14.6 seconds/token). A full installation requires approximately 1.7 TB of local disk space, while streaming mode needs only 215 GB but reduces inference speed to over 3 minutes/token. The project provides an OpenAI-compatible API server supporting chat and code completion, but it is recommended to set the client timeout to the hour level.
• Kimi K3 goes open source: 2.8T MoE model and technical report (2026-07-27): Kimi releases its strongest model, Kimi K3, a 2.8T-parameter MoE model with native visual understanding and a 1M token context window. The new architecture achieves a 2.5x intelligence improvement per unit of computation. In addition to model weights, Kimi also open-sources high-performance attention kernels, MoE communication libraries, and large-scale agent runtime environment infrastructure.
• Kimi K3 open-sources distributed agent environment AgentENV (2026-07-27): We have open-sourced AgentENV in collaboration with kvcache-ai. AgentENV is a distributed system for running agent environments at scale. Its components support the reinforcement learning training of the Kimi K3 agent, featuring fast snapshot, recovery, and branching capabilities, making it suitable for large-scale parallel agent workflows. Explore on GitHub: http://github.com/kvcache-ai/AgentEnv
• Kimi K3 open day: model weights, technical report and key Infra technologies released together (2026-07-27): Moon's Dark Side releases the 2.8 trillion parameter mixture-of-experts model Kimi K3, supporting native visual understanding and a 1 million token context window. Its scaling efficiency is 2.5 times higher than Kimi K2.5, and it simultaneously open-sources model weights, a technical report, and three Infra technologies: MoonEP, FlashKDA, and AgentEnv.
• Kimi releases visual perception benchmark PerceptionBench (2026-07-27): Kimi.ai released PerceptionBench, a visual perception benchmark derived from the failure patterns of current frontier models across 42 benchmarks. This benchmark deconstructs visual perception into 10 atomic capabilities and constructs 3,000 validation questions, each testing only a single perceptual ability without requiring reasoning or external knowledge.
• Kimi K3 goes live on the official site, built for agentic coding and knowledge work (2026-07-16): The official site has listed the Kimi K3 model (2026-07-16), positioned for agentic coding and knowledge work. The review at that time only mentioned K3 open-source commercial revenue-sharing terms and did not cover its specific features, use cases or newer research such as PerceptionBench.
• Kimi K3 Release (World's Largest Open-Source Model) (2026-07-16): 2.8 trillion parameters, world's largest open-weight model, KDA hybrid linear attention + 1 million token context + native visual understanding, #1 on Frontend Code Arena globally (1679 points surpassing Fable 5); official honestly states overall capabilities still lag behind Fable 5/GPT-5.6 Sol; complete weights open-sourced on July 27; July 19 reports Moonshot AI plans Hong Kong IPO within 6 months
• $31.5 Billion Valuation Financing + K2.7 Code (2026-06-30): $20 billion valuation financing closed, new round initiated at pre-money valuation of $31.5 billion; ARR exceeds $300 million; K2.7 Code programming model released, K3 planned for July release
• Kimi K2.6 (2026): Moonshot AI releases K2.6 (MIT open-source), 13-hour continuous coding + 300 Agent cluster; web/app default to K2.6
• Kimi K2 / K2.5 (2025): Transition from long-text assistant to Agent + coding direction
• Early Kimi Chat (2023-2024): One of the first domestic AI dialogue products featuring 2 million character ultra-long context