This Kimi Model Timeline tracks the major releases of Moonshot AI’s Kimi model family, from the early K1.5 reasoning model to the latest K3 generation.
Kimi started as a long-context AI assistant and later expanded into a model series used for general AI tasks, coding, and agent workflows.
Moonshot AI is the company behind Kimi. The Kimi name now covers the AI assistant, foundation models, coding products, and developer services, while individual model releases such as K2, K2.5, K2.6, and K3 mark the technical progression of the model family.
Kimi K3 is the latest major release in this timeline. It follows a series of models that gradually expanded Kimi’s reasoning, coding, multimodal, and agent capabilities. The timeline below records the releases that shaped Kimi’s development.
Last updated: July 27, 2026.
Latest Kimi Model Update
Kimi K3 launched on July 16, 2026 as Moonshot AI’s new general-purpose flagship. It replaced K2.6 as the recommended model for general Kimi tasks and became available through Kimi.com, Kimi Work, Kimi Code, and the Kimi API.
The model supports text and visual input, a one-million-token context window, coding, reasoning, and long multi-step work. Moonshot AI released the complete K3 weights eleven days after the hosted launch. Hosted access and downloadable weights therefore have separate dates in the timeline.
Planning to use Kimi through an API? Compare current Kimi model costs with GPT, Claude, Gemini, DeepSeek, and other providers in our AI LLM API pricing guide.
Kimi Model Timeline at a Glance
| Date | Event | What Changed |
|---|---|---|
| July 27, 2026 | Kimi K3 full weights | Moonshot AI completed the open-weight release after the hosted launch. |
| July 16, 2026 | Kimi K3 | The new general flagship added native vision and a 1M-token context window. |
| June 12, 2026 | Kimi K2.7 Code | A dedicated thinking-only coding branch entered Kimi Code and the API. |
| May 25, 2026 | Original K2 API series retired | Older K2 preview, turbo, and thinking model IDs lost support. |
| April 20, 2026 | Kimi K2.6 | The general model improved long coding tasks and agent workflows. |
| February 9, 2026 | Agent Swarm preview | K2.5 coordinated parallel sub-agents for larger tasks. |
| January 27, 2026 | Kimi K2.5 | Visual input joined reasoning, coding, and agent capabilities. |
| November 6, 2025 | Kimi K2 Thinking | Long-form reasoning and repeated tool use became a dedicated K2 mode. |
| September 5, 2025 | Kimi-K2-Instruct-0905 | An updated K2 checkpoint improved coding and expanded context to 256K. |
| July 11, 2025 | Kimi K2 | The first open-weight K2 models focused on coding, tools, and agents. |
| April–June 2025 | Specialist Kimi models | Vision, audio, mathematics, coding, and research branches appeared. |
| January 20, 2025 | Kimi K1.5 | The numbered model line began with multimodal reasoning research. |
| March 20, 2024 | Long-context expansion | Kimi expanded its prompt capacity to two million Chinese characters. |
| October 2023 | Kimi assistant testing | Moonshot AI introduced its long-context assistant to early users. |
2023: Kimi Starts as a Long-Context AI Assistant
Early 2023: Moonshot AI Is Founded
Moonshot AI was founded in Beijing in early 2023. The company concentrated on large language models, long-context systems, and the infrastructure needed to process large amounts of text in one session.
October 2023: Kimi Enters Testing
Kimi first appeared in October 2023 as an AI assistant built for long Chinese documents and extended conversations. Early users could upload or paste large amounts of material, ask questions about it, and request summaries without splitting the source into many short prompts.
2024: Long Context Defines the Kimi Product
March 20, 2024: Kimi Expands to Two Million Chinese Characters
Moonshot AI expanded Kimi’s supported prompt length to as many as two million Chinese characters. The larger limit made the assistant useful for books, research files, financial documents, legal material, and other sources that exceeded the context available in many competing chat products.
This period established the feature most closely associated with early Kimi: you could bring unusually long material into one conversation and continue asking questions without repeatedly rebuilding the context.
June 26, 2024: Mooncake Research Addresses Long-Context Serving
Moonshot AI published Mooncake, an architecture for serving long-context models more efficiently. Mooncake was not a new Kimi model, but it addressed the cost and speed problems created by large prompts and repeated context.
The research belongs in Kimi’s history because long-context products depend on both model capability and the systems that keep large sessions responsive.
2025: Kimi Becomes a Numbered Model Family
January 20, 2025: Kimi K1.5 Starts the Numbered Line
Kimi K1.5 marked the first major numbered model in Moonshot AI’s public research record. It focused on reasoning, reinforcement learning, long context, and tasks that combined text with images.
K1.5 changed how Kimi was presented. The brand was no longer defined only by a chatbot with a large context window. It now had a named model line with technical reports and model-level releases.
April–June 2025: Specialist Kimi Models Explore New Capabilities
Moonshot AI released several specialist projects between April and June 2025. They did not form a simple sequence between K1.5 and K2. Each branch explored a different capability that could later feed into general Kimi models.
- Kimi-VL / April 10: A vision-language model for images, video, documents, and visual agent tasks.
- Kimina-Prover Preview / April 15: A mathematical reasoning project focused on theorem proving.
- Kimi-Audio / April 26: An audio model for recognition, understanding, generation, and conversation.
- Kimi-Dev / June 17: A software-engineering model for repository-level coding work.
- Kimi-Researcher / June 20: A research agent for multi-step information gathering and synthesis.
July 11, 2025: Kimi K2 Establishes the Open-Weight Generation
Kimi K2 began the generation that brought Moonshot AI wider international attention. Moonshot AI released Kimi-K2-Base for adaptation and Kimi-K2-Instruct for chat, coding, tool use, and agent applications.
The model used a mixture-of-experts design with one trillion total parameters and 32 billion active for each token.
The first K2 Instruct release did not provide the dedicated long-thinking behavior that appeared later, and it did not support native visual input. Subsequent releases addressed both gaps.
September 5, 2025: Kimi-K2-Instruct-0905 Updates K2
The 0905 checkpoint improved agentic coding, instruction following, and context handling. Its supported context window increased to 256K, which gave coding and research agents more room for repositories, tool results, logs, and long instructions.
November 6, 2025: Kimi K2 Thinking Adds Deliberate Reasoning
Kimi K2 Thinking added a dedicated reasoning mode to the K2 family. It could work through longer chains of decisions and use tools repeatedly during research, coding, search, and writing tasks.
The model appeared in Kimi chat and through dedicated API endpoints. Moonshot AI later retired those original K2 Thinking model IDs after newer general models combined reasoning with vision and agent features.
2026: Kimi Combines Vision, Coding, and Agent Work
January 27, 2026: Kimi K2.5 Unifies Multimodal and Agent Features
Kimi K2.5 brought text, visual input, reasoning, coding, and agent workflows into one general model. Users could choose instant, thinking, agent, and Agent Swarm modes through Kimi’s hosted products.
The model also connected visual input with coding. You could provide a screenshot, video, or interface reference and ask Kimi to inspect or reproduce the design through code.
February 9, 2026: Agent Swarm Enters Research Preview
Moonshot AI presented Agent Swarm as a K2.5 research preview. One coordinating agent could divide a large task among many sub-agents and combine their work.
Agent Swarm showed how Kimi was moving from one assistant completing steps in sequence to a system that could distribute research and production work in parallel.
April 20, 2026: Kimi K2.6 Focuses on Longer Coding Tasks
Kimi K2.6 improved long-horizon software work, instruction following, and agent execution. It retained text and visual input, a 256K context window, and both thinking and non-thinking modes.
Moonshot AI released K2.6 through Kimi.com, the Kimi app, Kimi Code, the API, and open model artifacts. It remained the main general Kimi model until K3 arrived.
May 25, 2026: Original K2 API Models Lose Support
Moonshot AI discontinued the original K2 preview, turbo, and thinking API identifiers on May 25. This retirement did not remove every model with K2 in its name. K2.5 and K2.6 remained separate supported model families.
The change affects developers who find older sample code. An integration that references a retired K2 endpoint needs a supported model ID from the current Kimi Platform list.
June 12, 2026: Kimi K2.7 Code Creates a Coding Branch
Kimi K2.7 Code launched as a specialist model for long software-engineering tasks. It uses thinking mode by default and targets coding agents that need to follow instructions across large repositories and extended tool sessions.
Using Kimi for software development? Our Kimi Code CLI guide covers installation, commands, MCP servers, skills, plugins, providers, and browser-based sessions.
July 16, 2026: Kimi K3 Becomes the New General Flagship
Kimi K3 increased the model family to 2.8 trillion total parameters and expanded the context window to one million tokens. It also brought native visual understanding into the main flagship model.
K3 was designed for long coding sessions, research, reasoning, document work, and visual tasks. It launched through Kimi.com, Kimi Work, Kimi Code, and Kimi Platform. The same model supported consumer chat, workplace tasks, coding agents, and API applications.
Moonshot AI also introduced Kimi Delta Attention and Attention Residuals in K3. These are architecture changes intended to handle long sequences and move information through a larger model.
July 27, 2026: Full Kimi K3 Weights Become Available
Moonshot AI published the complete K3 model weights eleven days after the hosted release. The later date marks the open-weight milestone for organizations that wanted to evaluate K3 through compatible third-party infrastructure.
Current Kimi Models and Discontinued Versions
- Kimi K3: The current general flagship for visual input, long context, coding, reasoning, and knowledge work.
- Kimi K2.7 Code: A current specialist model for coding agents, with standard and high-speed hosted variants.
- Kimi K2.6: A supported general multimodal model with thinking and non-thinking modes.
- Kimi K2.5: A previous multimodal model. It is unavailable to newly registered platform users, with a full platform sunset scheduled for August 31, 2026.
- Original K2 API identifiers: The 0711 preview, 0905 preview, turbo, and K2 Thinking IDs were discontinued on May 25, 2026.
- Moonshot V1 models: Older context-length variants belong to the pre-K-series API era and are moving out of the modern platform lineup.
How the Kimi Model Family Changed Over Time
Kimi’s first advantage was the ability to read unusually long source material. The 2023 assistant and the 2024 context expansion were aimed at documents, research, and conversations that exceeded common prompt limits.
The focus shifted in 2025. K1.5 established a named model line, while the vision, audio, mathematics, coding, and research releases tested separate capabilities. K2 then placed coding, tool use, agents, and downloadable weights at the center of Moonshot AI’s international model strategy.
K2.5 and K2.6 merged those capabilities into general multimodal agent models. K3 continued the same direction with a much larger model, native vision, and a one-million-token context window. Kimi now connects one model family with chat, workplace software, coding tools, hosted APIs, and open-weight deployment.
Compare other fast-moving model families: Continue with the Qwen Model Timeline and the DeepSeek Timeline.
Kimi Model Timeline FAQ
Who develops the Kimi models?
Moonshot AI develops the Kimi model family and operates Kimi chat, Kimi Work, Kimi Code, and Kimi Platform. The company was founded in early 2023.
Is Kimi the same as Moonshot AI?
No. Moonshot AI is the company. Kimi is its assistant and product brand. K1.5, K2.6, K2.7 Code, and K3 are model names used across Kimi products.
Is Kimi K3 open source?
Kimi K3 is an open-weight model with published parameters and a Kimi license. Open weights do not automatically mean that the training data, training process, software, and redistribution terms match traditional open-source software. Review the model license before commercial redistribution or modification.
Can developers access Kimi models through an API?
Yes. Kimi Platform provides hosted API access to supported models. Model IDs, prices, context limits, and rate limits can change, so developers should confirm the current platform list before deployment.
Can Kimi models run locally?
Several Kimi releases have downloadable weights and support from third-party inference projects. Local deployment depends on the checkpoint, quantization, inference engine, and hardware.
What is the difference between Kimi K2.6, K2.7 Code, and K3?
K2.6 is a general multimodal model for chat, reasoning, coding, and agent work. K2.7 Code is a specialist thinking model for software engineering. K3 is the newer general flagship and has a one-million-token context window.
