📡 Breaking news
Analyzing latest trends...
AI Text-to-Speech.

Moonshot AI Drops Kimi K3 A 2.8T Multimodal Model with Instant Cloud Partner Availability.

Moonshot AI Drops Kimi K3 A 2.8T Multimodal Model with Instant Cloud Partner Availability.
Moonshot AI Releases Kimi K3 Open Weights Under Custom License with Instant Cloud Ecosystem Support

Moonshot AI has officially released its highly anticipated Kimi K3 open-weight model, delivering on its previous roadmap promises. The model is distributed under the proprietary Kimi K3 License, which grants open access and modification rights while establishing clear commercial and attribution thresholds for enterprise scale.

Under the license terms:

  • MaaS Providers: Model-as-a-Service (MaaS) hosting platforms generating over $20 million in annual revenue must enter into a formal commercial agreement with Moonshot AI before offering Kimi K3 hosting services.

  • Third-Party Applications: Developers and companies building downstream products on Kimi K3 can do so freely. However, services exceeding 100 million monthly active users (MAU) or $20 million in annual revenue are required to prominently display "Kimi K3" branding within their product interfaces.

Architecturally, Kimi K3 features a Mixture-of-Experts (MoE) design with 2.8 trillion total parameters and 104 billion active parameters (2.8T-A104B). It natively supports multimodal input (text and vision) and utilizes a hybrid quantization scheme combining MXFP4 and MXFP8 data formats to optimize inference efficiency and memory footprint.

To ensure immediate day-one enterprise accessibility, Moonshot AI launched the model alongside a initial cohort of cloud infrastructure partners, including DigitalOcean, Nebius, Fireworks, Baseten, and Modal. All launch partners are offering unified API access priced at $3.00 per million input tokens and $15.00 per million output tokens.

The way AI labs are redefining "open source"—instead of traditional open-source licenses (like Apache 2.0 or MIT)—they're adopting hybrid commercial licenses driven by revenue and user size. By setting a $20 million revenue threshold for MaaS providers and requiring explicit references for large-scale consumer applications, Moonshot AI protects its commercial rights from major cloud vendors while making its models accessible to open-source developers and early-stage startups.

As open-weighted models scale to trillions of inclusive parameters, rendering standard FP16 or FP8 requires very high GPU VRAM allocations. Microscaling models allow high-parameter MoE models to run with significantly reduced memory bandwidth without compromising output quality, making 2.8T parameter models practical for mid-sized enterprise use cases.

Historically, when open-weighted models were released, developers had to wait days or weeks for cloud providers to build the infrastructure to deliver and fine-tune their quantization kernels. By coordinating simultaneous launches across developer-centric cloud infrastructure (such as Fireworks, Baseten, and Modal) at a fixed price ($3/$15), Moonshot AI ensures immediate enterprise deployment and API parity from day one.

 

 

Source: HuggingFace 

💬 AI Content Assistant

Ask me anything about this article. No data is stored for your question.

Comments

Popular posts from this blog

Seagate launches HDD capacity, size 12TB

Romanian Land Registry Wiped by Hackers, Freezing Real Estate Transactions Nationwide.

OpenAI Admits Unreleased AI Model Escaped Sandbox to Hack Hugging Face During Cyber Test.

White House Accuses Moonshot AI of IP Theft via Covert Claude Fable Distillation and Thai Server Links.

AMD Advancing AI Helios Rack Launches with 72 GPUs Claiming 30% Edge Over NVIDIA Rubin.

Google Launches Video Selfie Verification for Secure, Deepfake-Resistant Account Recovery.

Anthropic Debuts Claude Opus 5: Fable 5-Level Intelligence at Half the Cost.