On July 27, 2026, Moonshot AI simultaneously released the Kimi K3 model weights and a technical report on Hugging Face and GitHub—eleven days after the model debuted as an API service.
K3's scale alone makes it newsworthy. It has 2.8 trillion total parameters and a 1-million-token context window. Actually running it, however, requires roughly 594 GB of storage, and Moonshot recommends a "super-node" configuration with 64 or more accelerators. This isn't something the average developer can casually spin up locally, but for organizations with cloud infrastructure or large-scale GPU environments, it offers a way to break free from dependence on proprietary APIs.
Along with the weights, Moonshot open-sourced high-performance attention kernels, an MoE communication library, and a set of tools serving as an agent foundation. Because Kimi Delta Attention (KDA) is incompatible with conventional prefix caching, the company directly contributed a KDA-compatible cache implementation to the vLLM project.
The Architecture Behind "the First Open Model in the 3-Trillion-Parameter Class"
K3 uses a Mixture-of-Experts (MoE) structure. Out of 896 total experts, only 16 are actually activated per token. Of the model's 93 layers, 69 are handled by "Kimi Delta Attention," while the remaining 24 layers are processed by gated latent-attention. A mechanism called "Attention Residuals" also selectively pulls representations across depth rather than stacking them uniformly.
Moonshot states that this combination of architectural choices, together with a routing method called Stable LatentMoE, achieved roughly 2.5x the scaling efficiency compared to Kimi K2. However, this figure comes from the company itself and has not been through independent reproduction.
There's another specification that could cause significant rework if overlooked during deployment. K3 was trained using what's called "preserved thinking history mode," which requires passing all prior assistant utterances, reasoning content, and tool calls into every subsequent input. Moonshot explicitly states that switching models mid-conversation or omitting this history will cause outputs to become unstable.
For the developer community, this specification changes the design premise. It's not an issue for use cases contained within a single conversation, but for long-running tasks or agent configurations orchestrating multiple models, inserting K3 partway through is difficult. You either need to run a K3-dedicated conversation thread from the very start of a session, or redesign an existing multi-model framework specifically for K3. One can read part of the motivation behind open-sourcing the inference infrastructure as an intent to solve these implementation hurdles together with the community.
The Revenue Line Beyond the MIT-Style Clause
The license design poses more of a question to the industry than the model itself does.
Read just the grant-of-rights section, and the Kimi K3 License is nearly MIT. Anyone can use, copy, modify, distribute, sublicense, sell, deploy, fine-tune, and create derivatives for free. However, two conditions attach depending on the scale of revenue generated.
① Commercial license requirement for resale businesses: MaaS resellers—businesses that provide third parties with access to inference or fine-tuning and give them control over inputs and parameters—must enter into a separate agreement with Moonshot once their annual revenue exceeds $20 million. A typical example would be a SaaS company earning revenue by reselling the Kimi K3 API.
② Attribution requirement for large-scale commercial products: Commercial products with 100 million or more monthly active users, or monthly revenue exceeding $20 million, are required to display "Kimi K3" in their UI.
Internal use on one's own infrastructure, or access via Moonshot's certified inference partners, is exempt from these conditions.
The dividing line is drawn not by user scale but by business model. It's designed so that a startup can freely use K3 for research and product development, and Moonshot's involvement only kicks in once that startup grows into a large-scale API resale business. This differs from the ecosystem that emerged around DeepSeek's MIT license, which placed no restrictions on commercial use. It's also a structure in which the more K3 spreads, the greater Moonshot's negotiating leverage becomes.
Cloud provider behavior will be a key point going forward. As of this reporting, none of the three major hyperscalers—AWS Bedrock, Microsoft Azure AI Foundry, or Google Vertex AI—have added Kimi K3 to their offerings. It appears each company is weighing the political risk of offering a large-scale open model originating from China as a managed service. This position cuts both ways for Moonshot. The absence of major clouds creates demand for specialized inference providers hosting K3, or for companies pursuing their own deployments. But conversely, a model that isn't on major clouds is harder to fit into corporate procurement processes, meaning adoption will likely take longer.
Evaluation Results and the Gaps That Remain
K3's benchmark standing is clear. Artificial Analysis measured K3's Intelligence Index score at 57 as of the API's release, placing it at roughly third overall, on par with Claude Opus 4.8 and GPT-5.5. It falls short of Claude Fable 5 and GPT-5.6 Sol, but stands out prominently among open-weight models. Given that GLM-5.2 scores 51 and DeepSeek V4 Pro scores 44, the gap is not small.
Agentic performance figures are even more striking. K3's Elo rating on GDPval-AA v2 is 1668, a substantial jump from its predecessor Kimi K2.6's 1190. On AutomationBench-AA, it ranks first at 53%, with a cost per task of $0.94—on par with GPT-5.6 Sol and roughly half the cost of Claude Opus 4.8.
On the other hand, independent testing by the UK Cyber Institute points to a significant gap versus frontier models in cyber capabilities. A similar gap has been observed in mathematical ability, and several researchers have suggested this may indicate skewed capabilities resulting from distillation. However, the evidence remains circumstantial at best.
Moonshot itself has announced scores of 93.5% on GPQA Diamond and 91.2 on BrowseComp, though its technical report notes that "evaluation harness conditions differ from model to model." With the weights now public, it has become possible for parties other than Moonshot to attempt reproducing these figures.
The distillation allegations and the benchmark gaps need to be considered separately. Even if some degree of distillation did occur, that alone may not fully explain the weaknesses in cyber and mathematical capability. It's possible that the composition of training data and differences in evaluation harnesses are interacting in complex ways, and determining exactly how much each factor contributes will require external researchers to gain direct access to the weights and run repeated experiments. Before the weights were released, that access simply didn't exist.
The Market Reacted First, Washington Followed
The market impact of K3's arrival was felt as early as the API release stage. Shares of Chinese AI competitors fell—Z.ai by as much as 30%, MiniMax by as much as 16%, and Alibaba by 4%. In blind arena evaluations, developers rated K3 higher than U.S. models on frontend coding tasks.
Moonshot's own numbers moved as well. Daily revenue increased at least sixfold following K3's release (per Bloomberg reporting). ARR grew in three leaps—$100 million in March 2026, $200 million in April, and $300 million in June 2026—and the company is now simultaneously pursuing a funding round at a $50 billion valuation and preparing for an IPO in Hong Kong.
The reaction in Washington was sharp. On July 22, 2026, White House Office of Science and Technology Policy (OSTP) Director Michael Kratsios publicly accused Moonshot of procuring NVIDIA GB300 chips in circumvention of export controls and of conducting large-scale distillation from Anthropic's Fable 5. Treasury Secretary Scott Bessent stated that entity-list designation and sanctions were "on the table," and reports indicate the Commerce Department's BIS has opened an investigation.
Moonshot has not issued an official comment on these allegations. Company employees have argued that with only 15 days between Fable 5's release and the K3 announcement, conducting large-scale distillation would have been physically impossible. Distillation itself is a widely used industry technique, and even among U.S. developers and open-weight advocates, viewing it as inherently improper is a minority position—but its application by a Chinese lab to a U.S. model is being treated as a separate issue. The release of the weights brings independent verification of the training process somewhat closer, but tracing training data remains far from easy.
NVIDIA and Microsoft jointly declared their support for open-weight AI on July 24, and on the same day OpenAI's Sam Altman traveled to Washington with plans for a GPT successor model in hand. Into that landscape arrived a 2.8-trillion-parameter model out of China, downloadable by anyone. This adds one more immovable fact to the AI regulation debate.
Well before BIS's investigation reaches any conclusion, the developer community's own attempts to reproduce K3's results and evaluate its capabilities will likely take the lead. Independent verification of its cyber capabilities, and benchmark reproducibility once the recommended harness is removed, will be the next things to watch.