Moonshot AI

Moonshot AI was founded in spring 2023 to pursue artificial general intelligence. Its stated approach emphasizes open models and user participation; its website lists a contact address on Zhichun Road in Beijing’s Haidian District.⁠[2]

Moonshot AI launched Kimi Chat. Founder Zhilin Yang subsequently described its roughly 200,000-Chinese-character context and multi-document question answering at WISE, also connecting the company’s name with Pink Floyd’s album The Dark Side of the Moon.⁠[3]

Alibaba’s fiscal 2025 annual report described its fiscal 2024 approximately US$800 million preferred-stock investment in Moonshot for about 36% equity. This historical transaction disclosure does not establish an unchanged ownership percentage in later years.⁠[4]

Mooncake research described Kimi’s serving platform, separating prompt prefill from decoding and using CPU, memory and SSD resources for disaggregated caching. Its scheduler balances throughput against latency targets for long-context and overloaded workloads.⁠[5]

The Kimi team published k1.5 research combining long-context scaling, policy optimization and multimodal training for reasoning. It also investigated transferring long-chain reasoning into shorter responses. A training report and evaluations do not themselves constitute a release of all model weights.⁠[6]

Kimi-VL used a mixture-of-experts language decoder with about 2.8 billion active parameters and MoonViT for native-resolution images. Its research covered image-text understanding, OCR, multiple images and long videos, with 128K context and publicly available code and models.⁠[7]

Kimi K2 introduced Base and Instruct versions with one trillion total and 32 billion active parameters. Training used MuonClip for stability, synthetic tool-use data and reinforcement learning. The original release did not yet support visual input.⁠[8]

The company introduced kimi-k2-turbo-preview, retaining the contemporary K2 model’s parameter scale and context settings while focusing on inference speed. Its announcement included time-limited launch promotions, which were not permanent pricing commitments.⁠[9]

K2 Thinking interleaved reasoning with tool use for search, coding and complex problems. The company demonstrated tasks involving 200–300 sequential tool calls and offered chat and API access. Those task lengths reflected its demonstrations and evaluations.⁠[10]

In an AMA recap, the founder and research team explained that K2 Thinking favored more reasoning tokens for quality, with INT4 quantization and the Turbo API improving speed. Vision remained under development, and serving million-token context was then considered expensive.⁠[11]

K2.5 continued training K2 on visual and textual data, enabling image- and video-assisted coding and visual debugging. Its Agent Swarm research preview used parallel-agent reinforcement learning: an orchestrator dynamically divided work across up to 100 subagents and 1,500 coordinated steps.⁠[12]

AnthropicAnthropicAnthropic is an artificial intelligence research and product company founded in 2021 that develops the Claude models, assistant, and tools such as Claude Code, and operates as a public benefit corporation. Dario Amodei and Daniela Amodei are its co-founders, and its research covers model safety, interpretability, and steerability.Open the full entry accused Moonshot of extracting Claude capabilities through over 3.4 million exchanges involving reasoning, tools and code. This was Anthropic’s allegation.⁠[13]

K2.6 targeted sustained coding and complex tasks; an official example described code optimization lasting over 12 hours. Agent swarms expanded to 300 subagents and 4,000 coordinated steps for documents, websites, slides and spreadsheets. Demonstrations do not guarantee autonomous success on every task.⁠[14]

K3 combined 2.8 trillion parameters, native vision and million-token context with Kimi Delta Attention and Attention Residuals. The company warned that missing thinking history or switching models mid-session could destabilize quality, and that excessive initiative required explicit behavioral boundaries.⁠[15]

The international policy named NOVASCENT PRIVATE LIMITED as provider and data controller, described service- and model-improvement uses of inputs, files and generated content, and provided access, correction and deletion channels. Enterprise training commitments are separately stated and should not be conflated with personal services.⁠[16]

Anthropic’s later safeguards article alleged that some user requests to Kimi had been secretly forwarded to Claude for distillation. This account came from the complainant; it was neither a judicial finding nor a description of every Kimi request.⁠[17]

K3.1 Pro increased post-training over K3, retaining million-token context and adding a Highspeed variant and 3D Make. The company said it removed some cybersecurity data during training to reduce misuse. Full weights were scheduled for November 4, still a future plan as of October 30.⁠[18]

Kimi Work supports Windows and Apple-silicon Macs, connecting local files, scripts and scheduled tasks with browser-extension workflows. Its product documentation says it requests authorization before modifying local files or running related code.⁠[19]

Kimi Code provides desktop, command-line and IDE-plugin interfaces for code search, feature implementation, fixes and refactoring, including tests, builds and scripts. These are coding-tool workflows, a product layer distinct from the underlying Kimi models.⁠[20]

Kimi Business provides separate team workspaces and seat management, with a company commitment not to train on enterprise data. The usage-based Kimi API offers project budgets, balance alerts and rate limits; business seats and API calls represent different service models.⁠[21]