Moonshot AI Releases Kimi K3 as First Open 3T-Class Frontier Model

Kimi K3 is a 2.8 trillion parameter native multimodal model with a 1 million token context window and advanced attention mechanisms released by Moonshot AI.

Moonshot AI announced the release of Kimi K3 on July 16, 2026. The model represents a significant step in the company's efforts to compete in the frontier model space. It is described as the most capable model from the company to date. The 2.8 trillion parameters place it in the 3T class. Native multimodal capabilities allow it to handle vision tasks alongside text. The 1 million token context window enables processing of very long documents and conversations. Availability through the Kimi platform allows immediate use for developers and researchers. Kimi Work targets enterprise knowledge work. Kimi Code focuses on software engineering tasks. The API provides programmatic access for integration into other systems. Performance is reported to be near that of leading U.S. models like GPT-5.6 and Fable 5 on agentic coding benchmarks. The company emphasizes its design for long-horizon tasks and self-evolving workflows. This release comes as part of a broader push by Chinese AI firms to develop open frontier models.

The development builds on previous iterations in the Kimi series. Kimi K2 served as the prior version with lower scaling efficiency. The new model achieves an approximate 2.5 times improvement in overall scaling efficiency. This comes from the architectural changes including Kimi Delta Attention and Attention Residuals. The company has positioned the model for use in deep reasoning scenarios. Knowledge work applications include analysis of large datasets. Software engineering benefits from the long context for understanding entire codebases. The open weights release scheduled for July 27, 2026, will allow the community to build upon the model. This openness contrasts with some closed models from other providers. Moonshot AI aims to foster innovation through this approach. The Stable LatentMoE is part of the design for efficient expert activation. Only 16 experts are activated out of 896 in the mixture of experts setup. This sparse activation helps maintain efficiency despite the large parameter count.

What background context surrounds the Kimi K3 launch?

Moonshot AI has been active in the AI space with a focus on large language models. The company previously released models that paved the way for this frontier effort. The competitive landscape includes U.S. companies developing similar scale models. Kimi K3 is presented as matching top systems on specific tasks. Agentic coding involves the model acting as an autonomous agent in coding projects. Long context supports maintaining state over extended interactions. The timing of the release on July 16, 2026, allows for early adoption before the open weights. The announcement includes details on the technical innovations. These innovations address challenges in training and inference at this scale. Training efficiency is improved by about 25 percent with the Attention Residuals. The additional cost is less than 2 percent according to the company. This efficiency gain is significant for large scale training runs. The hybrid linear attention in Kimi Delta Attention speeds up decoding substantially.

The model is designed to handle self-evolving workflows where the system can improve its own performance over time. This capability is important for complex projects that require iterative refinement. In software engineering, the long context window allows the model to consider large repositories of code at once. This reduces the need for chunking and maintains coherence. For knowledge work, the multimodal aspect enables understanding of charts and images in addition to text. Deep reasoning tasks benefit from the large parameter count and efficient architecture. The company claims the model reaches performance levels comparable to the most advanced closed models. Benchmarks are reported near those of GPT-5.6 and Fable 5. The open nature of the upcoming weights release could accelerate research in the field. Community contributions may lead to further improvements and fine tunes. The release marks a milestone in open frontier model development from China.

What are the key technical specifications of Kimi K3?

Kimi K3 features 2.8 trillion parameters in a native multimodal setup. The context window extends to one million tokens. This allows for extensive input without truncation in most cases. The architecture incorporates Kimi Delta Attention which is a hybrid linear attention mechanism. This design contributes to faster inference times. Attention Residuals are included to boost training efficiency. The model uses a mixture of experts approach with 896 total experts. During inference, it activates only 16 of these experts. This selective activation reduces computational overhead. Native vision capabilities are built in for image understanding. The model supports tasks requiring both text and visual inputs. The API documentation confirms the parameter count and context size. These specifications position the model for demanding applications. The combination of scale and efficiency is a key differentiator.

Key architectural features of Kimi K3

Architecture Element	Description	Reported Benefit
Kimi Delta Attention	Hybrid linear attention mechanism	Up to 6.3x faster decoding in million-token contexts
Attention Residuals	Innovative residual connections	~25% higher training efficiency at less than 2% additional cost
MoE Expert Activation	Activates 16 out of 896 experts	Efficient computation at frontier scale
Scaling Efficiency	Overall architectural improvements	2.5x improvement compared to Kimi K2

The technical details highlight the focus on efficiency at scale. Kimi Delta Attention allows for up to 6.3 times faster decoding when dealing with million token contexts. This is crucial for practical use in long context scenarios. Attention Residuals deliver about 25 percent higher training efficiency. The cost increase is minimal at under 2 percent. These features make training and running the model more feasible. The MoE design with limited active experts keeps inference costs manageable. The 2.5 times scaling efficiency improvement over the previous model indicates better utilization of compute resources. This is important as models grow larger. The native multimodal support expands the range of possible applications. Vision integration is seamless within the same model. The company has provided documentation on the API for developers to experiment with these features.

Today, we are introducing Kimi K3 — our most capable model. Kimi K3 is a 2.8T-parameter model built on our Kimi Delta Attention and Attention Residuals, with native vision capabilities and a 1-million-token context window. It is the world's first open 3T-class model, designed for frontier intelligence across long-horizon coding, knowledge work, and reasoning.Moonshot AI / Kimi team

How does Kimi K3 impact the market and stakeholders?

The release of Kimi K3 has implications for the AI market particularly in the open model segment. By promising open weights, Moonshot AI provides an alternative to closed frontier models. Developers and researchers gain access to a large scale model without restrictions. This could lead to increased innovation in agentic systems and coding assistants. Enterprises may adopt it for internal knowledge work tools through the Work platform. The focus on long horizon tasks opens new possibilities for automated software development. Stakeholders in the U.S. may see this as increased competition from Chinese AI efforts. The performance matching top systems on specific benchmarks adds to the competitive pressure. Open weights allow for customization and fine tuning by the community. This democratizes access to frontier capabilities. The scheduled release date for weights gives time for initial closed testing. Overall the move strengthens the position of open source in high end AI.

Market dynamics may shift as more open models of this scale become available. Companies relying on proprietary models might face pressure to offer similar openness. The efficiency improvements could lower the barrier for running such models on available hardware. Self-evolving workflows could transform how AI is used in production environments. Knowledge work automation sees potential gains from the long context and multimodal features. The API access allows integration into existing workflows quickly. This immediate availability supports rapid adoption. The company targets software engineering and deep reasoning as primary use cases. These areas have high demand for advanced AI assistance. The announcement has drawn attention to Moonshot AI as a key player in frontier models.

What are the expert reactions and future outlook for Kimi K3?

The company announcement describes Kimi K3 as the most capable model to date. It highlights the design for frontier intelligence. The open 3T class status is emphasized as a first. Reactions from the broader community are anticipated following the open weights release. The technical advancements in attention mechanisms are likely to be studied closely. Other AI labs may incorporate similar ideas in their designs. The performance on agentic coding tasks positions it as a strong contender. Future developments could include further optimizations and applications. The July 27, 2026, date for open weights is a key milestone. Until then, access is through the provided platforms. This phased approach allows for feedback and refinement. The model is expected to influence the direction of open frontier model development.

What platforms provide access to Kimi K3 currently?

1. Kimi platform at kimi.com
2. Kimi Work for enterprise use
3. Kimi Code for software engineering
4. Kimi API for programmatic integration

The availability across multiple platforms ensures broad accessibility. Users can choose the interface that fits their needs. The API supports custom applications and integrations. This flexibility is important for different stakeholder groups. The model is ready for use in production environments where appropriate. The open weights will expand this further once released. The current setup allows testing and evaluation ahead of the full open release. This strategy helps gather user feedback. The platforms are designed to showcase the model's strengths in coding and reasoning. Overall the launch strategy supports both immediate use and long term openness.

The focus on self-evolving workflows represents an advanced use case for frontier models. These workflows allow the AI to adapt and improve based on task outcomes. In practice this could lead to more autonomous AI agents. The long context supports maintaining complex state information over time. This is essential for multi step reasoning processes. The native multimodal capabilities add another layer of functionality. Images and diagrams can be analyzed in context with text. This is useful in many professional domains. The efficiency gains make running the model more practical. Training at this scale benefits from the 25 percent efficiency boost. The overall design reflects careful engineering to balance performance and cost.

Competition in the frontier model space is intensifying with this release. Other companies will likely respond with their own advancements. The open weights approach may encourage similar moves from competitors. The Chinese AI ecosystem gains a prominent example of large scale open model. This could attract talent and investment to the sector. Researchers worldwide will have the opportunity to experiment with the model after July 27, 2026. The benchmarks near GPT-5.6 levels indicate high capability. Continued development is expected as the community engages with the open model. The initial closed period allows the company to monitor usage and performance. This data can inform future iterations. The release is a notable event in the 2026 AI landscape.