Geo News
Community curated by people like you
LatestAICryptoHealthWorld AffairsUS Politics
Moonshot AI Releases Kimi K3 as World's Largest Open-Weight Model at 2.8 Trillion Parameters
00

Moonshot AI Releases Kimi K3 as World's Largest Open-Weight Model at 2.8 Trillion Parameters

Jul 27, 2026

Moonshot AI has released the open weights for Kimi K3, a massive 2.8-trillion-parameter Mixture-of-Experts model, under the Apache 2.0 license. While the model is free to download, running its 1.4-terabyte file locally requires enterprise-grade hardware, such as at least eight H100 GPUs. To bypass these hardware barriers, cloud platforms Together AI and Modal have launched day-0 hosted access supporting Kimi K3's full 1-million-token context window. The release marks a major milestone in open-source AI, despite viral social media claims misrepresenting the ten-day launch timeline.

Kimi K3 open-weight release

  • ▪The full download of the Kimi K3 model weights is approximately 1.4 terabytes, while the MXFP4 quantized version is about 594 gigabytes
  • ▪Moonshot AI released the open weights for its Kimi K3 model on Hugging Face under the Apache 2.0 license
  • ▪Running the full Kimi K3 model locally requires at least eight H100-class GPUs or a single node of eight H200 GPUs

Model specifications

  • ▪Kimi K3 took first place on the Arena Frontend Code leaderboard with 1,679 points, ahead of Claude Fable 5 at 1,631 and GPT-5.6 Sol at 1,618
  • ▪Kimi K3 features 2.8 trillion total parameters and a 1,048,576-token context window

MoE architecture details

  • ▪Kimi K3 incorporates Kimi Delta Attention, a sparse attention mechanism designed to reduce memory costs and address the attention sink problem
  • ▪Kimi K3 is a Mixture-of-Experts model that utilizes 896 experts and activates 16 experts per token

Together AI hosting

  • ▪Together AI's day-0 hosting of Kimi K3 supports the model's full 1-million-token context window
  • ▪Together AI announced day-0 hosted access for Kimi K3, offering serverless API endpoints and dedicated GPU deployments

Modal hosting

  • ▪Modal announced day-0 hosted access for Kimi K3, providing cold-start optimizations and containerized vLLM templates

Timeline claims scrutiny

  • ▪Social media claims that Kimi K3 went from initial launch to open weights in under 24 hours are inaccurate
  • ▪The actual timeline from Kimi K3's initial API launch on July 16, 2026, to its open-weights release on July 26, 2026, was approximately ten days

4 sources

Findskill
Kimi K3 Is 'Free' Now — Here's Why You Still Can't Run It
View source article
Dev
Kimi K3 Open Weights Are Here: How to Self-Host the 2.8T-Parameter Model (Hardware, vLLM, and Data Sovereignty)
View source article
Startupfortune
Moonshot AI releases Kimi K3 open weights making the world's largest open-weight model free to download - Startup Fortune
View source article
Explainx
Kimi K3 Open Weights: 2.8T Params, Day-0 Hosting (2026) | explainx.ai Blog
View source article

Featured stories

View more in China tech & industrial strategy

DeepSeek releases V4.1-Flash AI model with drastically reduced costs and memory usage

Sep 10, 2026 · 12 sources

Chinese AI startup DeepSeek taps CITIC Securities for Shanghai IPO preparation

Sep 9, 2026 · 5 sources

IFM releases K2 Horizon family of six open models from 0.9B to 375B parameters

Sep 7, 2026 · 2 sources

Nvidia CEO Jensen Huang says AGI has arrived, congratulates OpenAI on GPT-6 Astra

Sep 6, 2026 · 4 sources

Share your thoughts

Will open-weight models eventually surpass closed commercial models?

Story comments

Loading comments…

Share your thoughts

Will open-weight models eventually surpass closed commercial models?

Related entities

China

Related Projects

Moonshot AI

Topics

China tech & industrial strategyAI safety & social impactLarge language models (LLMs)Open model deployment toolsOpen weights vs closed modelsOpen-source AIAI foundation models

Featured stories

View more in China tech & industrial strategy

DeepSeek releases V4.1-Flash AI model with drastically reduced costs and memory usage

Sep 10, 2026 · 12 sources

Chinese AI startup DeepSeek taps CITIC Securities for Shanghai IPO preparation

Sep 9, 2026 · 5 sources

IFM releases K2 Horizon family of six open models from 0.9B to 375B parameters

Sep 7, 2026 · 2 sources

Nvidia CEO Jensen Huang says AGI has arrived, congratulates OpenAI on GPT-6 Astra

Sep 6, 2026 · 4 sources