Kimi K3: Moonshot AI Releases 2.8T Open-Weight Model With 1M Context

Kimi K3 is a 2.8 trillion parameter open-weight mixture-of-experts model from Moonshot AI with native multimodal capabilities and a 1 million token context window.

Moonshot AI released the full weights of Kimi K3 around July 27, 2026. The company published the weights on Hugging Face along with a technical report and blog post. The release occurred under the Kimi K3 License which follows a modified MIT-style framework. This structure permits broad commercial and research use while retaining specific attribution requirements.

What background at Moonshot AI preceded the Kimi K3 launch?

Moonshot AI built Kimi K3 on a series of earlier Kimi models that focused on language and reasoning tasks. Those prior releases established baseline performance in agentic scenarios and coding assistance. The company identified scaling limitations in previous architectures that prompted the new design choices. Iterative development emphasized efficiency gains alongside parameter growth.

The progression reflects a consistent strategy of increasing model scale while introducing targeted architectural refinements. Earlier versions lacked the extensive context window and multimodal integration seen in Kimi K3. Moonshot AI documented these advancements in the accompanying technical blog. The timeline aligns with industry trends toward larger open-weight systems.

What technical specifications define the Kimi K3 architecture?

Kimi K3 operates as a mixture-of-experts model containing 896 total experts. The system activates 16 experts for each processed token. This configuration reduces active compute while maintaining capacity. Stable LatentMoE contributes to training stability across the large parameter count.

The vision component relies on the MoonViT-V2 encoder for native multimodal processing. This encoder handles images, charts, and screenshots directly within the model pipeline. The 1 million token context window enables extended document analysis and multi-turn agent interactions. Kimi Delta Attention combined with Attention Residuals forms the core attention mechanism.

These architectural elements yield approximately 2.5 times the scaling efficiency compared to Kimi K2. The efficiency metric measures intelligence output relative to compute investment. Full implementation details reside in the technical report published alongside the weights. The design prioritizes agentic workloads such as coding and tool orchestration.

What benchmark performance does Kimi K3 demonstrate?

Kimi K3 records 93.5 percent on the GPQA Diamond benchmark at maximum reasoning effort. The score appears in the official model card hosted on Hugging Face. This result reflects performance on graduate-level science questions that test deep reasoning.

The model further achieves 91.2 percent on BrowseComp at maximum effort. Terminal-Bench 2.1 reaches 88.3 percent when evaluated with the Kimi Code harness at maximum settings. These figures originate from the same Hugging Face model documentation.

Benchmark results reported for Kimi K3 in the Hugging Face model card.

Benchmark	Score	Evaluation Setting
GPQA Diamond	93.5%	Maximum reasoning effort
BrowseComp	91.2%	Maximum effort
Terminal-Bench 2.1	88.3%	Kimi Code harness, maximum

What access and deployment requirements apply to Kimi K3?

Users obtain the weights directly from the Hugging Face repository at moonshotai/Kimi-K3. The download includes the full parameter set and associated configuration files. Moonshot AI also provides API access through its platform for those preferring managed inference.

Self-hosting demands roughly 1.4 terabytes of storage capacity for the complete weights. Enterprise-grade GPUs are required to run inference at practical speeds. The modified MIT-style license governs redistribution and derivative work.

What steps enable initial deployment of the model weights?

1. Navigate to the Hugging Face repository moonshotai/Kimi-K3 and review the model card.
2. Download the full set of model weights to local or cloud storage meeting the 1.4 terabyte requirement.
3. Install compatible inference frameworks and verify hardware compatibility with enterprise GPUs.
4. Load the model using the provided configuration and test basic inference on sample inputs.
5. Consult the technical blog for optimization guidance on Kimi Delta Attention and multimodal features.

What market implications follow from the open-weight release?

The availability of Kimi K3 weights places direct competitive pressure on proprietary frontier model providers. Organizations specializing in coding agents and long-context applications now have a high-performance open alternative. The license terms encourage commercial integration and fine-tuning by resource-equipped teams.

Developers in the agentic AI sector gain the ability to run the model locally or on private infrastructure. This reduces reliance on external APIs for sensitive workloads. Multimodal and extended context features support new use cases in document processing and visual reasoning.

The release contributes to the broader trend of open models narrowing performance gaps with closed systems. Stakeholders can now evaluate and adapt a 2.8 trillion parameter system without licensing restrictions beyond the modified MIT framework. Moonshot AI positions the model as a foundation for further community-driven advancements.

What official reactions accompanied the Kimi K3 announcement?

Moonshot AI described Kimi K3 as its most capable model to date in the release materials. The statements emphasized native visual understanding alongside the expanded context window. Links to the weights, technical report, and blog accompanied the announcement.

Releasing the model weights and technical report of Kimi K3. Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window. New model architecture: 2.5x the intelligence per unit of compute, not just more params.Kimi.ai, Official account of Moonshot AI

A separate company statement reinforced the open-weight and native multimodal characteristics. The description highlighted the Kimi Delta Attention and Attention Residuals components. These communications directed users to the Hugging Face repository for immediate access.

What developments are anticipated after the Kimi K3 release?

Community fine-tunes and efficiency optimizations are expected to emerge rapidly due to the open weights. Integration into existing agent frameworks will likely accelerate given the multimodal and long-context strengths. Moonshot AI may issue additional documentation or API updates in response to usage patterns.

Further architectural refinements could appear in subsequent models from the company. The current release establishes a new baseline for open frontier systems in agentic benchmarks. Observers will track adoption metrics and downstream applications over the coming quarters.