Moonshot AI

Moonshot AI Large Model Open Platform provides long text data processing APIs based on Moonshot AI's large models. The open platform supports flexible API calls, delivering a leading technical experience.

AWS announced that Moonshot AI's Kimi K3 is now generally available on Amazon Bedrock, expanding enterprise access to the 2.8T-parameter flagship model. This marks a major hosted-platform milestone, giving AWS customers direct access to K3's multimodal capabilities through the Bedrock managed service. The integration allows developers and enterprises to invoke Kimi K3 alongside other frontier models in Bedrock without self-hosting, benefiting from AWS-managed scaling and compliance. Combined with Oracle OCI Enterprise AI support, K3 is now distributed across multiple top-tier cloud providers, signaling broad ecosyst

Oracle's September 2026 AI update confirmed that Moonshot AI's Kimi K3 is now available for import in OCI Enterprise AI as a multimodal model supporting both image and text inputs. This positions K3 for enterprise multimodal and agentic AI use cases inside Oracle's managed AI platform, complementing its existing AWS Be Alongside K3, OCI Enterprise AI expanded its Model Import catalog with offerings from AI Singapore, Mistral, Google, Alibaba, NVIDIA, DeepSeek, Z.ai, and Xiaomi — including SEA-LION v4 32B IT, Mistral Magistral Small, Gemma 4, Qwen 3.5/3.8, NVIDIA Nemotron 3 Super and 3.5 Lightning, DeepSeek V4 Pro and V4 Flash 0731, G

Moonshot AI's developer-tooling activity accelerated through September 2026, with Kimi CLI 1.51.0 archived on Sep 21 in favor of a new Kimi Code CLI. A follow-up update on Sep 22 routed existing CLI entry points to the Kimi Code installer, consolidating the user experience around a single command-line surface. Earlier Earlier releases in the chain reflect ongoing CLI maturation: v1.49.0 (Jul 16) stopped sending implicit reasoning effort and corrected the completion-budget calculation to use remaining context, while v1.48.0 (Jun 22) and v1.47.0 (Jun 5) delivered targeted fixes around tool-error rendering and tool-output inclusion. To

vLLM published a preview of production-scale Kimi K3 support developed jointly with Moonshot AI, NVIDIA, AMD, and the broader community in preparation for the open-weight release by July 27, 2026. The post outlines that vLLM is preparing model implementation, Docker images, deployment recipes, and production validation The release branch integrates FlashKDA, fused KDA decode, fused KDA projections and convolution, fused AttnRes, a reimplemented MLA module, SiTU-enabled MXFP4 MoE execution, and optimized expert routing, alongside multimodal integration and tool-calling parsers. NVIDIA-specific kernels are under final tuning while an i

Simon Willison's independent review of Kimi K3 corroborates Moonshot's announcement, confirming the 2.8-trillion-parameter scale and Moonshot's claim of the first "open 3T-class model," surpassing DeepSeek's 1.6T v4 Pro. He reports pricing of $3 per million input tokens and $15 per million output tokens — a sharp incre Citing the Artificial Analysis report, the review notes Kimi K3 reaches an Elo of 1547 on a private long-horizon knowledge-work evaluation (+732 over K2.6, behind only Claude Fable 5), at $0.94 per task versus $1.04 for GPT-5.6 Sol and $1.80 for Opus 4.8, while using 21% fewer output tokens than K2.6 on the Artificial

Kimi K2.7-Code, Moonshot AI's coding-focused open-weight model, was released on Hugging Face on June 12, 2026 under a Modified MIT license. The model is a 1-trillion-parameter Mixture-of-Experts with 32 billion active parameters, 384 experts, and a 256K context window. Moonshot reports K2.7-Code uses roughly 30% fewer On July 1, 2026, Kimi K2.7-Code became the first open-weight model available in GitHub Copilot's model picker, a milestone for enterprise adoption. GitHub announced general availability on July 12, 2026, with Copilot now routing across five model providers (OpenAI, Anthropic, Google, Microsoft, and Moonshot AI) under a

Moonshot AI releases Kimi-K2.6 model with 1T parameters, attention optimizations - SiliconANGLE