Moonshot

Moonshot AI is the Beijing organization behind Kimi, its language and vision models, API platform and Kimi Code. Its work connects model training with the infrastructure and tool use needed to complete tasks: stabilizing large-model optimization, serving long contexts efficiently, and training agents to coordinate work. By July 2026, its Kimi K3 model combined native vision with a million-token context window and 2.8 trillion total parameters.[1][3][7][6]

Moonshot dates its founding to spring 2023 and lists an address in Beijing’s Haidian District. The 2026 ZGC Forum profile describes Yang Zhilin as Kimi’s founder and CEO and says the Kimi app and website serve tens of millions of professionals and knowledge workers monthly.[1][14]

Earlier serving and optimizer research supplies context for the later Kimi models. Moonshot’s research chronology includes Mooncake in 2024 and scaling Muon in 2025, addressing how to serve long inputs and train larger models.[2][7][8]

Mooncake separates prefill, which processes input tokens to prepare an answer, from decoding, which generates subsequent tokens one at a time. These stages have different computational demands. Its serving architecture organizes separate clusters around the KV cache—stored attention state that avoids recomputing previously processed tokens—and uses CPU memory and SSD resources for distributed cache storage. Scheduling balances cache reuse and throughput against time to the first token and time between output tokens. Under overload, prediction-based early rejection avoids spending resources on requests that cannot finish within latency objectives.[7]

The Mooncake authors describe it as Kimi’s serving platform. Their reported experiments used replayed workload traces with a dummy model following LLaMA2-70B’s architecture. Relative to their baseline while meeting latency objectives, they report 75% more requests handled in replayed real-workload traces and up to a 525% throughput increase in certain simulated scenarios.[7]ⁿ¹

Muon is Scalable for LLM Training adapts an existing optimizer based on matrix orthogonalization to larger language models. The authors add weight decay and adjust update scale, introduce the Moonlight mixture-of-experts model, and report roughly twice the computational efficiency of AdamW in their compute-optimal scaling experiments. Moonshot’s contribution was scaling and adapting Muon; the paper credits the original optimizer to Jordan and collaborators.[8]ⁿ¹

During Kimi K2 development, the team encountered exploding attention logits, the scores that determine which tokens receive attention. It reports that existing approaches, including logit soft-capping and query-key normalization, were inadequate. Its response, MuonClip, rescales query and key projection weights after optimizer updates to control attention scores at their source. Moonshot reports pretraining K2 on 15.5 trillion tokens without a training spike using this optimizer.[3]

Released in July 2025, K2 had one trillion total parameters and 32 billion activated parameters: its mixture-of-experts architecture uses only part of the parameter pool for each token. Moonshot offered Base for adaptation and fine-tuning and Instruct for chat and agentic use. Tool training combined simulated environments, rubric-based filtering and reinforcement learning with a self-judging critic. Verifiable tasks helped update the critic used to assess less readily verifiable work, such as research reports. At launch, K2 lacked vision support; the developers also disclosed excessive token generation, incomplete tool calls and degraded performance on some tool-use tasks and whole-project prompts.[3][2]

Kimi K2.5, announced in January 2026, added native multimodality through continued pretraining on approximately 15 trillion mixed visual and text tokens. Images and videos could guide code generation, while screenshots of rendered results could support visual debugging. Moonshot demonstrated reconstructing a website from video and iterating on an interface using visual inspection. Its accompanying coding product, Kimi Code, works in the terminal, integrates with editors including VSCode, Cursor and Zed, and supports image and video inputs.[4][2]

K2.5 also introduced an Agent Swarm research preview advertising up to 100 subagents. Parallel-Agent Reinforcement Learning, or PARL, trains an orchestrator to divide tasks among dynamically instantiated subagents whose models remain frozen. To counter serial collapse, where the orchestrator works sequentially despite having parallel capacity, training initially rewards agent creation and subtask completion. Those auxiliary rewards taper to zero as training shifts toward overall task success. Rewarding completed subtasks also discourages spawning workers merely to inflate parallelism metrics.[4]

The swarm’s Critical Steps metric counts orchestration work and the slowest concurrent branch at each stage. It measures the path that holds up completion, rather than total computation across workers: additional agents help only when they shorten that path. Moonshot reported up to 4.5× lower wall-clock time in wide-search scenarios compared with single-agent execution; the benefit depends on the workload and its decomposition.[4]

Announced in July 2026, Kimi K3 increased model size to 2.8 trillion total parameters, with 104 billion activated, and combined native vision with a million-token context window. Kimi Delta Attention addresses information flow across long sequences, while Attention Residuals selectively retrieves representations across model depth. Its technical abstract also describes million-token agentic reinforcement learning with persistent rollout and sandbox states. Moonshot acknowledges that K3’s overall performance trails Claude Fable 5 and GPT-5.6 Sol, with a noticeable user-experience gap relative to both.[5][6][2]ⁿ¹

K3’s developers disclose two operational limitations. Generation can become unstable when the software managing conversations and tools omits historical thinking content or switches an ongoing session from another model to K3; they recommend a compatible harness such as Kimi Code and avoiding mid-session switching. K3 can also make unexpected decisions under ambiguous intent or when encountering minor issues. Applications requiring defined behavioral boundaries need explicit constraints in the system prompt or AGENTS.md.[5]

Kimi’s model-selection documentation, observed in September 2026, lists K2.6 for fast Q&A, K3 for complex conversations and editable documents, and K3 Cluster for large-scale parallel work. Chat and Agent have merged, with reasoning-effort choices controlling the depth and token use of reasoning. Consumer access to K3’s million-token context requires the highest-tier membership benefit. The documentation advertises up to 300 coordinated subagents and more than 4,000 parallel tool calls for the cluster; these are product capacity claims.[9]

In a February 23, 2026 account, competing model provider Anthropic alleged that Moonshot generated over 3.4 million exchanges with Claude to extract capabilities in agentic reasoning, tool use, coding, data analysis, computer use and vision. Anthropic said it attributed the campaign through request metadata matched to public profiles of senior Moonshot staff.[10]ⁿ²

Anthropic’s September 2026 report further alleged that Moonshot silently forwarded Kimi customer requests to Claude, displayed Claude’s responses to users and saved some exchanges for model training. It described almost 300,000 relayed requests in one ten-day instance, separately from over 23 million observed exchanges attributed to Moonshot between May and July 2026. Its examples included internal code and live credentials submitted by an engineer, and targeted-person CCTV data from Chengdu submitted by a user Anthropic assessed as likely PLA-affiliated. That uncertain assessment concerns the user, not Moonshot conducting surveillance. Anthropic said it did not know whether Moonshot notified customers that their requests were being rerouted.[11]ⁿ²