Kimi K3 Open Weights Released: Moonshot Calls It the First Open 3T-Class Model, at 2.8 Trillion Parameters

Chinese AI startup Moonshot AI has released the full weights of its large language model Kimi K3 on Hugging Face¹. When the company announced the model on July 16, it committed to publishing the full weights “by July 27, 2026”³, and it met that deadline. At 2.8 trillion total parameters, Moonshot AI positions Kimi K3 as the first open model to reach 2.8 trillion parameters and the world’s first open model in the 3-trillion-parameter class.

The published repository (huggingface.co/moonshotai/Kimi-K3) contains the model card and MXFP4-format weights, available for anyone to download¹. The developer community’s response has been strong, with the release becoming a major topic of discussion on Hacker News.

Model Specs: 896-Expert MoE and a 1M-Token Context Window

According to the model card, only 104B of Kimi K3’s 2.8 trillion total parameters are activated at inference time¹. The model scales up Mixture of Experts (MoE) sparsity with a framework called Stable LatentMoE, activating just 16 of its 896 experts¹. MoE is a technique that keeps compute costs down by selectively running only parts of a huge model — see our explainer on Mixture of Experts (MoE) for details.

The attention mechanism builds on the company’s own Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), combining 69 KDA layers with 24 Gated MLA layers¹. The context length is 1,048,576 tokens (1 million), and the model has native vision (image understanding) capabilities¹.

Evaluations listed on the model card as of July 23, 2026 show scores of 93.5 on GPQA Diamond, 67.5 on the coding benchmark DeepSWE, 88.3 on Terminal-Bench 2.1, and 91.2 on the web-research benchmark BrowseComp¹. Details of the architecture, training, and evaluations were slated to be published alongside a technical report³.

License: MIT-Based, but Large-Scale Commercial Use Has Conditions

The weights were released under a custom license called the “Kimi K3 License”¹. It is a modified version of the MIT license: research, learning, and small-scale use are essentially unrestricted, while large-scale commercial deployment comes with conditions².

Specifically, if a licensee and its affiliates generate more than $20 million in aggregate revenue over any consecutive 12 months from Model-as-a-Service offerings built on Kimi K3 (such as providing the model via API), a separate agreement with Moonshot AI is required². In addition, commercial services with more than 100 million monthly active users or more than $20 million in monthly revenue must display “Kimi K3” in their UI².

For companies considering building the model into their products, the revenue clause is the practical dividing line. Internal use and small-to-mid-scale services face no real constraints, while cloud providers reselling the model and large AI services will need to negotiate a contract.

The Operational Barrier: About 1.4TB Even at 4-Bit

Kimi K3 was built with quantization-aware training (QAT), which bakes quantization into the training process; the published weights are in MXFP4 (4-bit floating point) format with MXFP8 activations¹. MXFP4 is a 4-bit format with per-block scaling factors, natively supported by NVIDIA’s Blackwell and AMD’s MI400-generation accelerators⁴.

Even so, the scale of 2.8 trillion parameters is substantial: storing the weights alone requires roughly 1.4TB (about 5.6TB in FP16)⁴⁵. Loading that into inference memory implies something on the order of eighteen 80GB-class accelerators⁵. This is not a model an individual or small team can run locally — “self-hosting” is realistic only for organizations with serious GPU infrastructure.

The Backstory: Released 11 Days After Launch, Amid Sanctions Warnings

Kimi K3 launched on July 16 via API and consumer apps, with the weight release promised at that time (see our coverage of the announcement)³. Moonshot AI open-sourced the 1-trillion-parameter Kimi K2 in July 2025, meaning the company has grown its open-model scale 2.8x in a single year.

The release is also notable for landing in the middle of escalating US-China tensions. On July 22, White House OSTP Director Kratsios alleged on X that Moonshot had distilled Anthropic’s Fable without authorization to develop K3, and Treasury Secretary Bessent raised the possibility of sanctions and Entity List designation (see our article on the distillation allegations). Five days after being called out by name by government officials, Moonshot released the weights without walking back its commitment.

The Self-Hosting Ceiling Rises, but the API Is the Realistic Path

For readers evaluating AI adoption, this release means the performance ceiling of open-weight models has reached the 3T class. For industries that cannot let data leave their own infrastructure, having a near-frontier model available for self-hosting matters. Combined with recent releases like Thinking Machines’ open model “Inkling” from the company led by former OpenAI CTO Mira Murati, the gap between open and proprietary models keeps narrowing.

That said, running the model takes on the order of eighteen accelerators, so for most companies the realistic path remains the API or a hosting provider. And with the US administration raising the possibility of sanctions over the distillation allegations, evaluation criteria need to include the provider’s geopolitical risk and license terms — not just performance and cost. One more factor for model selection: now that the weights are public, third-party verification and auditing of the model has become possible.