Moonshot AI Releases Kimi K3 Open-Weight Model with 2.8 Trillion Parameters

Moonshot AI Drops Full Weights for Kimi K3

Moonshot AI released the full weights for Kimi K3, making the 2.8 trillion parameter model available for developers to download, modify, and self-host1

1

3

3

. The Chinese AI startup's move completes the rollout that began earlier in July when Kimi K3 debuted through Moonshot's hosted API. The open-weight model is now available on Hugging Face, where it reached No. 1 on the platform's trending chart within 30 minutes, recording more than 4,000 likes in that timeframe—the fastest growth Hugging Face had ever recorded5

5

. By the next day, the repository had garnered approximately 7,700 likes and tens of thousands of downloads.

The release includes the complete 2.8 trillion parameter Mixture-of-Experts model, inference infrastructure, optimized attention kernels, MoE communication libraries, and deployment components aimed at researchers and enterprise developers2

2

. Moonshot also provided a 47-page technical report documenting training innovations including Kimi Delta Attention, Attention Residuals, and Stable LatentMoE. The model activates 104 billion parameters from a pool of 896 experts while supporting native multimodal reasoning and a 1-million-token context window2

2

.

Kimi K3 Performance Rivals Leading Closed Models

Kimi K3 has demonstrated performance comparable to GPT-5.6 Sol and Claude Fable 5 across multiple benchmarks. The model scored 88.3 on Terminal-Bench 2.1, compared with 88.8 for GPT-5.6 Sol and 88.0 for Claude Fable 51

1

. It also outperformed both on Moonshot's reported BrowseComp results. Arena blind evaluations showed developers selecting Kimi K3 ahead of leading U.S. models on front-end coding tasks4

4

.

The most powerful open-source model of the year excels particularly in agentic workflows. It can understand existing projects, make changes across multiple files, and verify whether the final result works without requiring constant intervention1

1

. Real-world testing confirms the model performs at approximately 90% of GPT-5.6 Sol's capability across tasks like building landing pages, writing code, and fixing bugs.

API pricing makes Kimi K3 especially competitive. The model costs $3 per million uncached input tokens and $15 per million output tokens, while cached input drops to $0.301

1

. By comparison, GPT-5.6 Sol costs $5 and $30, Claude Opus 5 costs $5 and $25, and Claude Fable 5 costs $10 and $50 for the same input and output token counts.

Self-Hosting Requires Data-Center-Level Infrastructure

While the downloadable AI model is freely available, running Kimi K3 independently demands substantial resources. The official minimum requirement is eight B300/GB300-class GPUs or eight AMD MI355X GPUs, providing roughly 2.3TB of combined GPU memory1

1

. An Nvidia DGX B300 weighs approximately 370 pounds, occupies 10 rack units, contains eight 288 GB GPUs, and draws up to 14.5 kW. It also produces almost 49,500 BTU of heat every hour, requiring just over four tons of continuous cooling.

Setting up the infrastructure requires dedicated 200-240V power with multiple high-current circuits, continuous cooling, and an isolated location capable of handling data-center fans1

1

. Eight-GPU B300 servers currently start at around $470,000, with other configurations exceeding $500,000. Including taxes, shipping, suitable rack, UPS equipment, cooling, and electrical work, a realistic project budget would be approximately $550,000-$700,000. The mixture-of-experts design activates 104 billion parameters at a time, while the MXFP4 weights alone are expected to occupy roughly 1.4TB before accounting for runtime overhead3

3

.

Kimi K3 Usage License Includes Revenue-Based Restrictions

The Kimi K3 usage license grants broad rights to developers and enterprises to download, modify, and deploy the model for commercial purposes, but imposes obligations on larger companies and AI service providers not found in traditional open-source licenses2

2

. If the licensee or any of its affiliates operates a Model as a Service business, and the aggregate revenue of the licensee and its affiliates exceeds $20 million in total over any consecutive 12 months, the licensee must enter into a separate agreement with Moonshot AI before using the software for any commercial purpose.

Additionally, if the software is used for commercial products or services that have more than 100 million monthly active users, or more than $20 million in monthly revenue, "Kimi K3" must be prominently displayed on the user interface2

2

. These requirements do not apply to internal use or any use accessed through Moonshot AI's official products or certified inference partners.

Release Triggers US Policy Debate and Market Impact

The release arrives amid growing US policy debate over Chinese AI and open-weight models. White House Office of Science and Technology Policy Director Michael Kratsios accused Moonshot of training Kimi K3 using banned Nvidia chips and conducting large-scale model distillation against U.S. models, including Anthropic's Fable model3

3

4

4

. Moonshot AI has not responded to requests for comment on the allegations.

Meanwhile, Nvidia CEO Jensen Huang is backing a push to protect open-weight AI. The number of companies supporting a letter published on July 24 reportedly doubled from 25 to 50 within a day, adding names such as OpenAI, Google, AMD, Cisco, Cloudflare, GitHub, Block, and Ollama3

3

. The group is urging Washington not to place sweeping restrictions on downloadable models.

Kimi K3's debut sent Chinese AI competitor stocks sharply lower. Baidu stock lost as much as 30% of its value in Hong Kong trading, MiniMax Group stock fell as much as 16%, and Alibaba stock dropped 4%4

4

. Since K3 launched, Moonshot's daily revenue has grown by a factor of at least six. The company reached $300 million in annual recurring revenue in June, up from $200 million in April, and is seeking a new funding round at a $50 billion valuation ahead of a potential Hong Kong IPO as soon as this year.