Moonshot AI released the full model weights for Kimi K3 on Monday, opening direct access to a 2.8-trillion-parameter artificial intelligence system designed for coding, multimodal reasoning and long-duration agentic work. The publication on Hugging Face completes the second stage of a launch that began on July 16, when Moonshot made the model available through its Kimi consumer products, coding tools and hosted application programming interface.

The weight release is significant because it allows organizations to operate Kimi K3 outside Moonshot’s managed services. Developers can download the model, integrate it with supported inference frameworks, adapt it to specialized tasks and deploy it within privately controlled infrastructure. That creates a materially different commercial proposition from closed frontier models, whose parameters remain inaccessible and whose capabilities generally must be consumed through vendor applications or APIs.

Moonshot describes Kimi K3 as the first openly released model in the three-trillion-parameter class. Its exact total is 2.8 trillion parameters, but it uses a mixture-of-experts architecture that avoids activating the entire network for every token. The model card lists approximately 104 billion activated parameters, 896 routed experts, 16 selected experts per token and two shared experts. The approach is intended to preserve the capacity of an extremely large model while containing the computational cost of individual inference requests.

The architecture consists of 93 layers, including 69 layers using Kimi Delta Attention and 24 using gated multi-head latent attention. Moonshot says Kimi Delta Attention provides a more efficient foundation for processing long sequences, while its Attention Residuals system selectively retrieves representations from different depths of the network rather than relying only on conventional sequential accumulation. The company combines those components with a Stable LatentMoE framework intended to improve routing and training stability at high levels of sparsity.

Moonshot claims the resulting architecture delivers an approximately 2.5-fold improvement in overall scaling efficiency compared with Kimi K2. Kimi K2, released in 2025, contained one trillion total parameters and activated 32 billion. Kimi K3 therefore represents a substantial expansion in both overall model capacity and the amount of computation engaged during inference, even as its sparse design keeps active parameter counts far below the headline total.

The model’s context window extends to 1,048,576 tokens, allowing it to process large software repositories, extensive collections of documents or long-running agent histories in a single working context. Kimi K3 also includes MoonViT-V2, a 401-million-parameter vision encoder, providing native support for text and image inputs. Moonshot’s hosted implementation additionally supports video workflows, although the downloadable model card formally identifies text and images as the released model’s modalities.

The model is targeted particularly at long-horizon coding and knowledge-work tasks. Moonshot says Kimi K3 can navigate large repositories, operate terminal tools and continue engineering workflows with limited human intervention. The company has presented examples involving GPU-kernel optimization, compiler development, scientific computing, chip-design experiments and the production of interactive research reports.

Those demonstrations are company-produced and should not be treated as independent verification of production reliability. They nevertheless show the market segment Moonshot is pursuing: complex agentic workloads in which a model must gather information, use multiple tools, retain extensive context and complete a sequence of interdependent actions rather than answer a single prompt.

Moonshot’s published benchmark table places Kimi K3 close to leading proprietary systems across selected reasoning, programming and agent evaluations. The model card reports scores of 93.5 on GPQA Diamond, 88.3 on Terminal-Bench 2.1, 91.2 on BrowseComp and 94.5 on MCPMark-Verified. It also reports competitive results on software-engineering evaluations including FrontierSWE, ProgramBench, DeepSWE and SWE-Marathon.

Benchmark comparisons require caution because different models were sometimes paired with different agent frameworks, reasoning settings or fallback systems. Moonshot evaluated Kimi K3 using combinations of Kimi Code and other harnesses, while competing models were often tested with their respective coding environments. Several scores were taken from external leaderboards, while others came from Moonshot’s own runs or internal benchmarks. The company itself acknowledges that Kimi K3’s overall experience still trails the strongest proprietary models despite reaching frontier-level performance on parts of its evaluation suite.

The open-weight release makes independent testing more practical. Research laboratories and infrastructure providers can now evaluate Kimi K3 using consistent prompts, identical hardware assumptions and standardized agent frameworks. Enterprises can also test issues that public leaderboards rarely capture, including latency, output stability, domain-specific accuracy, observability, security controls and the operational cost of serving sustained workloads.

Deployment will remain difficult for all but well-funded organizations. Kimi K3 uses quantization-aware training with MXFP4 weights and MXFP8 activations, reducing memory and computational demands compared with higher-precision representations. Even so, a 2.8-trillion-parameter model requires data-center-scale memory, networking and accelerator capacity. Moonshot recommends supernode configurations with 64 or more accelerators, reflecting the need for a large, high-bandwidth communication domain when distributing experts across machines.

That requirement limits the practical meaning of “open” for individual developers. The files can be downloaded, but operating the full model efficiently is not comparable with running a smaller open-weight system on a workstation. Universities, cloud platforms, sovereign computing programs and large enterprises are more likely to become direct operators, while smaller developers will probably access Kimi K3 through hosted inference providers or quantized derivatives.

The model card provides integration paths for Hugging Face Transformers, vLLM, SGLang and Docker Model Runner. It also points developers toward third-party quantizations intended for llama.cpp, Ollama, LM Studio and compatible applications, although extremely aggressive compression would still leave substantial hardware requirements. Together AI was listed as an inference provider shortly after the release, indicating that commercial hosting partners are beginning to package the model for customers that do not want to manage its infrastructure directly.

Moonshot has also contributed support for Kimi Delta Attention prefix caching to the vLLM ecosystem. Prefix caching can reduce the cost of repeatedly processing common prompt material, an important feature for coding agents and enterprise applications that reuse large system instructions, repositories or document collections. Moonshot says its official service achieves a cache-hit rate exceeding 90% on coding workloads.

The company prices its hosted API at $0.30 per million cache-hit input tokens, $3 per million cache-miss input tokens and $15 per million output tokens. Open weights give customers an alternative to those usage charges, but self-hosting substitutes infrastructure, engineering and maintenance expenses for API fees. The economic decision will depend on workload volume, accelerator utilization, data-governance requirements and the extent to which a company intends to customize the model.

Enterprises may be attracted by the ability to keep prompts, documents and generated outputs inside private networks. Financial institutions, technology companies and government organizations frequently restrict the transmission of sensitive data to external AI services. Running Kimi K3 in a controlled environment could address some of those concerns, although adopters would still need to assess the model’s training provenance, behavior, software dependencies, cybersecurity exposure and applicable regulatory requirements.

The Kimi K3 License broadly permits users to copy, modify, publish, distribute, sublicense, sell, deploy and fine-tune the software and model weights. It is nevertheless a bespoke license rather than a standard permissive software license without commercial thresholds. A licensee operating a model-as-a-service business must enter into a separate agreement with Moonshot before using Kimi K3 commercially when the licensee and its affiliates generate more than $20 million in aggregate revenue during any consecutive 12-month period.

The license also requires prominent display of the “Kimi K3” name when the model or a derivative is used in a commercial product exceeding 100 million monthly active users or $20 million in monthly revenue. Those provisions do not apply to internal use or to access through Moonshot’s official products and certified inference partners. The conditions give researchers, startups and most enterprise users broad latitude while preserving Moonshot’s ability to negotiate with large inference platforms and mass-market technology companies.

The distinction between open-weight and open-source remains important. Publishing trained parameters allows extensive deployment and modification, but it does not necessarily provide everything required to reproduce the model from the beginning. Training datasets, complete data-curation processes and all internal development systems are not automatically disclosed with the weights. Kimi K3 therefore offers far more control than a closed API while falling short of full reproducibility under the traditional software meaning of open source.

For the broader AI market, the release strengthens the position of Chinese laboratories in the global open-model ecosystem. Alibaba’s Qwen family, DeepSeek’s systems and earlier Moonshot models have already attracted international developers seeking lower costs and greater control. Kimi K3 pushes that strategy closer to the frontier by pairing open weights with a parameter count, context capacity and agentic feature set more commonly associated with the most expensive proprietary platforms.

The release may place additional pressure on U.S. model developers to offer more capable downloadable systems. Closed providers retain advantages in integrated services, safety infrastructure, reliability, distribution and access to large computing fleets. Open-weight competitors, however, can gain adoption through customization, private deployment and integration into third-party tools that do not generate recurring revenue for the original developer.

Kimi K3 could also generate demand elsewhere in the technology supply chain. Operating a sparse model of this size requires advanced accelerators, high-bandwidth memory, fast interconnects, distributed inference software and specialized cloud capacity. The model may therefore create opportunities for chipmakers, hosting providers and optimization companies even when Moonshot is not directly collecting an API fee.

Near-term adoption will depend less on the 2.8-trillion-parameter headline than on whether infrastructure providers can make the model economical and dependable. Quantizations, optimized kernels, expert-parallel serving systems and compatible agent frameworks will determine achievable throughput and latency. Independent evaluations will also need to establish whether Kimi K3’s benchmark strength translates into consistent performance across real enterprise workloads.

Moonshot has disclosed several limitations. The model was trained to preserve its reasoning history, and performance can become unstable when an agent framework fails to return the complete prior assistant state or switches to Kimi K3 in the middle of a session. The company also warns that the model may behave too proactively when faced with ambiguous instructions, making explicit system-level boundaries important for applications in which autonomous actions could create operational or financial consequences.

The July 27 publication turns Kimi K3 from a hosted product with a promised future release into an independently deployable platform. Its immense hardware footprint will prevent broad local use, and its commercial license will require careful review by major service providers. Even with those constraints, the availability of frontier-scale weights gives developers a new alternative to proprietary AI and raises the competitive stakes for laboratories seeking to control the software, infrastructure and economics of advanced model deployment.