What Is Kimi K3?

The global AI landscape shifted measurably on July 16, 2026. Moonshot AI, a Chinese startup founded just three years ago, released KIMI K3, a 2.8 trillion parameter model that immediately claimed the title of the world's largest open-weight AI system ever built. In the same week, it posted coding benchmark scores that surpass Claude Fable 5 and placed fourth overall among all frontier models globally, behind only Claude Fable 5 and GPT-5.6 Sol on general capability while beating Claude Opus 4.8 and GPT-5.5 in head-to-head comparison.

This is not a regional story. KIMI K3 is a global frontier event. It changes what open-weight AI means, what Chinese AI labs have achieved, and what any organization or practitioner working with frontier models needs to understand about the options now available to them.

For technology professionals who want to develop the structured technical literacy needed to evaluate frontier models like KIMI K3, understand their architectural innovations, and make informed deployment decisions, building that foundation through a recognized Deep Tech Certification covering AI systems, model architectures, and enterprise deployment gives practitioners the grounding to engage with these developments analytically rather than reactively.

This guide answers the question completely: what is KIMI K3, how does it work, what does it benchmark, how is it priced, and why does it matter.

What Is KIMI K3?

KIMI K3 is Moonshot AI's flagship large language model and the most capable AI system the company has ever released. It launched on July 16, 2026, through the Kimi app, the Kimi Code platform, and the Kimi API. Full open-weight model files are scheduled for public release by July 27, 2026 under a Modified MIT license.

At its core, KIMI K3 is a Mixture-of-Experts large language model with approximately 2.8 trillion total parameters. It is the third major generation in Moonshot's Kimi model family, following the original Kimi chatbot released in 2023 and the K2 series that progressed through K2, K2.5, K2.6, and K2.7 Code between 2025 and early 2026.

The simplest way to understand KIMI K3's significance: it is the first model in AI history to exceed three trillion parameters at the open-weight tier. The previous record holder, DeepSeek V4 Pro, had 1.6 trillion parameters. KIMI K3 at 2.8 trillion is 75% larger. And Moonshot is releasing the weights openly, meaning any organization with the compute infrastructure to run it can download, host, and fine-tune it freely.

Why Was KIMI K3 Built?

Moonshot AI closed a $500 million Series C round in January 2026 at a $4.3 billion valuation. The funding was explicitly earmarked for KIMI K3 development and for expanding the compute infrastructure needed to train and serve a 2.8 trillion parameter model. The launch was previewed a day early when a promotional page on Moonshot's own Kimi Open Platform leaked before the scheduled announcement, generating immediate attention across the global AI research and developer community.

The strategic objective is visible in Moonshot's own positioning language. At launch, the company described KIMI K3 as "a new frontier of intelligence," acknowledging that it still trails Claude Fable 5 and GPT-5.6 Sol on overall capability while claiming it "consistently outperforms other tested models" across its evaluation suite. That framing reflects a specific ambition: to operate at the global frontier tier, not at a distance from it, and to do so with an open-weight model that any organization can access and adapt.

Bloomberg summarized the broader context: KIMI K3 is "the latest sign that Asian AI labs are closing a technology gap with the US," with the model outperforming all rivals except Anthropic's Claude Fable 5 and OpenAI's GPT-5.6.

KIMI K3 Architecture: How It Works

Mixture-of-Experts With 896 Experts

KIMI K3 uses a Mixture-of-Experts architecture with 896 experts and 16 experts activated per token during inference. MoE models do not apply all parameters to every token. Instead, a routing mechanism selects the most relevant subset of experts for each specific piece of input, making frontier-scale parameter counts economically viable by reducing the active compute per inference call compared to dense models at the same parameter size.

This architecture is why KIMI K3 can offer 2.8 trillion parameters at a price point of $3 per million input tokens. Dense models at a fraction of that parameter count would be significantly more expensive to serve per token.

Kimi Delta Attention (KDA): The Core Innovation

The architectural feature that distinguishes KIMI K3 most sharply from prior models is the Kimi Delta Attention mechanism. KDA is a hybrid linear-attention system that combines standard full attention with linear attention residuals.

Standard transformer attention has quadratic memory complexity with respect to sequence length, which means that scaling to one-million-token context windows using pure standard attention is extremely expensive. KDA addresses this directly: by combining linear attention with delta-residual corrections, the architecture achieves near-full-attention quality at a fraction of the memory cost for very long sequences.

This is the technical foundation for KIMI K3's one-million-token context window. The architecture makes the context length feasible without the prohibitive infrastructure cost that standard attention would require at that scale.

Always-On Thinking Mode

Where the K2 series required a separate reasoning variant for step-by-step deliberation, KIMI K3 unifies the reasoning capability into its default behavior. Thinking mode runs on every interaction. The model reasons through problems before generating a final response, matching the direction of the broader frontier where always-on reasoning has become the expected baseline rather than an optional mode.

Native Vision and Multimodality

KIMI K3 processes images natively, through its core model architecture rather than a separately attached vision module. This integration means that tasks combining visual and textual reasoning, such as chart analysis, scientific diagram interpretation, technical document understanding, and multimodal research synthesis, benefit from the full depth of the model's 2.8 trillion parameters applied jointly across text and image, rather than having the image processed by a lighter-weight appended system.

KIMI K3 Variants: K3 Max and K3 Swarm Max

Two variants shipped at the July 16 launch:

K3 Max is the primary general-purpose variant. It is the model powering the Kimi app chat interface, the Kimi Code coding assistant, and the default API endpoint. K3 Max is optimized for conversational tasks, knowledge work, complex coding, long-document analysis, and multi-step agentic workflows.

K3 Swarm Max is a second variant purpose-built for large-scale parallel processing. It is designed for environments where many simultaneous agentic workloads need to run concurrently at high throughput, such as large-scale data processing pipelines, batch inference jobs, and parallel multi-agent orchestration systems.

KIMI K3 Benchmark Performance

Moonshot AI published benchmark comparisons at launch, and independent evaluations from Artificial Analysis and individual researchers have corroborated the broad competitive picture.

Global Frontier Ranking

On independent overall evaluation, KIMI K3 ranks fourth among all frontier models globally. The ranking order from the top: Claude Fable 5, GPT-5.6 Sol, then KIMI K3, ahead of Claude Opus 4.8.

Coding: KIMI K3 Leads the Frontier

The result that received the most attention at launch is K3's coding benchmark performance. On coding-specific evaluations, KIMI K3 surpasses Claude Fable 5, making it the strongest available model for software engineering tasks at any tier, proprietary or open-weight. On SWE-bench Verified, which measures multi-file software engineering capability, K3's score positions it among the top three models globally.

The Kimi Code platform confirmed this with live user feedback at launch: early adopters describe K3 as competitive with or superior to top-tier closed coding models for practical software development work.

GPQA: Expert-Level Scientific Reasoning

On GPQA, which evaluates expert-level reasoning across biology, chemistry, and physics, KIMI K3 outperforms GPT-5.5 and is competitive with Claude Opus 4.8, while trailing Claude Fable 5 and GPT-5.6 Sol.

Long-Context Benchmarks

Given K3's one-million-token context window, Moonshot tested long-context retrieval and reasoning directly. K3 maintains strong performance across very long documents and multi-document retrieval tasks, confirming that the KDA architecture delivers practical long-context quality rather than nominal context window support that degrades in practice.

Human Preference Evaluation

Moonshot published pleasantness and conversational quality evaluations alongside technical benchmarks. K3 scored 75.5 on Moonshot's internal pleasantness metric, described as substantially ahead of prior Kimi models. This dimension matters for consumer-facing applications and enterprise interfaces where response quality and naturalness affect user adoption.

KIMI K3 Pricing: API and Consumer Access

Standard API Pricing

The cached input pricing is the most strategically significant figure. At $0.30 per million cached input tokens, KIMI K3's effective cost for code-heavy workflows that reuse the same context, such as a large codebase that serves as persistent context across many queries, is approximately one-tenth of the standard input price. Moonshot's Mooncake serving infrastructure is designed to maintain cache hit rates above 90% for coding workloads, which means the real-world effective input cost for software engineering applications is approximately four times lower than the nominal per-token rate.

For AI and enterprise professionals who want to develop the applied expertise needed to evaluate model cost structures like this and understand what they mean for production AI deployment decisions, a specialized Certified Artificial Intelligence (AI) Expert certification builds the foundational knowledge of AI systems, model architectures, and deployment economics that translates directly into better-informed technology decisions in fast-moving environments like the one KIMI K3 has created.

Consumer App Access

In the Kimi consumer app, K3 Max access starts at a ¥199 subscription tier. The app is available on iOS and Android, with Kimi Code available separately for software development workflows.

Launch Promotion

A recharge promotion offering 10% to 30% bonus API credits is running from July 15 through August 11, 2026, with the bonus percentage varying by recharge volume. For teams evaluating KIMI K3 for production integration, the promotion window provides a cost-effective opportunity for high-volume testing.

The Open-Weight Release: What July 27 Means

The full model weights for KIMI K3 are scheduled for public release by July 27, 2026 under a Modified MIT license. This is the single most consequential dimension of the launch for the global research and enterprise community.

What Becomes Possible After July 27

Self-hosting: Organizations can run KIMI K3 entirely on their own infrastructure, eliminating API dependency for privacy-sensitive applications, latency-critical deployments, or use cases requiring cost predictability at extreme scale.

Fine-tuning: KIMI K3 can be adapted to specialized domains using proprietary data. Healthcare, legal, financial, and industrial applications that require frontier-tier intelligence with domain specialization become directly feasible without sharing data with a third-party API.

Research access: The global research community gains access to a 2.8 trillion parameter MoE architecture built on the KDA attention mechanism for the first time, enabling reproducibility studies, architectural analysis, and second-order innovation built on the K3 foundation.

Derivative models: The Modified MIT license is commercially permissive. Organizations and developers can build, modify, and distribute derivative models based on K3 weights with relatively few restrictions.

No other 3-trillion-class model has been released at open-weight tier before July 27, 2026. When the weights land, KIMI K3 will represent a structural expansion of who can access frontier-tier AI capability.

KIMI K3 vs. Competing Frontier Models

KIMI K3 vs. Claude Fable 5

Claude Fable 5 leads KIMI K3 on overall benchmark performance. However, on coding benchmarks specifically, KIMI K3 surpasses Fable 5, and it does so at lower API pricing as an open-weight model. For software engineering-focused teams, the practical choice between the two depends on which benchmark category dominates their actual use case.

KIMI K3 vs. GPT-5.6 Sol

GPT-5.6 Sol leads on agentic coding benchmarks (Terminal-Bench 2.1: Sol at 88.8%) and on overall frontier evaluation. However, Sol is proprietary and closed-weight. KIMI K3 offers comparable performance across many dimensions as an open-weight model that can be self-hosted and fine-tuned, at lower pricing than Sol.

KIMI K3 vs. Claude Opus 4.8

KIMI K3 outperforms Claude Opus 4.8 on the majority of benchmark categories. This is the most actionable comparison for enterprise teams: Opus 4.8 is the current production default for many organizations, and K3's ability to exceed it as an open-weight model changes the cost-performance calculation materially.

KIMI K3 vs. DeepSeek V4 Pro

KIMI K3 at 2.8 trillion parameters is approximately 75% larger than DeepSeek V4 Pro at 1.6 trillion and substantially outperforms it on benchmark evaluations. KIMI K3 takes the title of largest open-weight model from DeepSeek.

What KIMI K3 Means for Different Audiences

For Software Engineering Teams

KIMI K3's coding benchmark leadership over Claude Fable 5, combined with the cached input pricing structure and Mooncake's 90%+ cache hit rate for code-heavy workflows, makes it a compelling production option for teams running high-volume software engineering workloads. The economics and the performance case both point in the same direction.

For Research Institutions

The July 27 open-weight release under a Modified MIT license makes KIMI K3 the most capable model available for academic and non-commercial research that requires self-hosting. The KDA architecture is itself a research contribution worth studying independently.

For Enterprise AI Teams

The combination of frontier-tier performance, open weights, self-hosting capability, and API-level pricing below Claude Opus 4.8 makes KIMI K3 a structurally different option than any prior open-weight model. Enterprise teams previously constrained to proprietary API-only models for frontier-tier tasks now have a genuinely viable alternative.

For AI Practitioners and Strategists

KIMI K3 demonstrates conclusively that the assumption of geographic concentration of frontier AI capability is no longer valid. Operating with an accurate picture of the global frontier model landscape now requires including Moonshot AI and KIMI K3 alongside Anthropic, OpenAI, and Google.

Limitations and Honest Caveats

Trailing on general capability: Despite leading on coding, KIMI K3 trails Claude Fable 5 and GPT-5.6 Sol on overall benchmark suites. Teams should run task-specific evaluations rather than assuming performance generalizes from any single benchmark category.

Weights not yet available: The July 27 open-weight release is scheduled but not yet complete as of July 16-18. Organizations planning around self-hosting or fine-tuning need to plan around that date.

Language coverage: K3's training and evaluation data are strongest in Chinese and English. Performance in other languages should be independently validated before committing to production deployment in non-Chinese, non-English markets.

No video or screen sharing: The Kimi app does not support real-time video or screen-sharing interaction with K3 at launch, unlike some competing multimodal platforms.

Strategic Communication for the KIMI K3 Era

As KIMI K3 reshapes the global frontier AI landscape, technology professionals increasingly need to communicate these developments clearly and credibly to stakeholders who are not following the technical details daily. Being able to explain what a 2.8 trillion parameter open-weight model means for enterprise AI strategy, how it changes vendor options, and why the open-weight release on July 27 matters for data privacy and cost control requires the same strategic communication skills that apply to any high-stakes technology positioning conversation. A Marketing Certification develops exactly these capabilities: the business positioning, stakeholder communication, and strategic framing skills that allow technology professionals to lead their organizations through developments like the KIMI K3 launch with clarity and confidence, alongside a Deep Tech Certification and a Certified Artificial Intelligence (AI) Expert credential.

Conclusion

KIMI K3 is the world's largest open-weight AI model, the first 3-trillion-class open-weight system, and the first open-weight model to lead coding benchmarks over proprietary frontier competitors. Released July 16, 2026, by Moonshot AI, it brings together a novel Kimi Delta Attention architecture, a one-million-token context window, native vision, always-on thinking mode, and frontier-tier benchmark performance at pricing below Claude Opus 4.8 and GPT-5.6 Sol. The July 27 open-weight release under a Modified MIT license will make it directly accessible for self-hosting and fine-tuning by any organization with the infrastructure to deploy it.

For practitioners and organizations working with frontier AI in 2026, KIMI K3 is not optional knowledge. It is the new baseline for what open-weight models can do and the clearest evidence yet that the frontier is global. Building the expertise to evaluate it well, through a Deep Tech Certification that grounds technical literacy in AI systems and deployment, a Certified Artificial Intelligence (AI) Expert credential that develops applied AI understanding, and a Marketing Certification that develops the strategic communication skills to lead these conversations within organizations, positions professionals to operate at the frontier with genuine competence and credibility.

Frequently Asked Questions

1. What is KIMI K3?

KIMI K3 is Moonshot AI's flagship large language model, released July 16, 2026. It is the world's largest open-weight AI model at 2.8 trillion parameters, featuring a Mixture-of-Experts architecture, a one-million-token context window, native vision, and always-on thinking mode.

2. Who built KIMI K3?

KIMI K3 was developed by Moonshot AI, a Chinese AI startup founded in March 2023 and backed by $500 million in Series C funding closed in January 2026 at a $4.3 billion valuation.

3. When was KIMI K3 released?

KIMI K3 launched on July 16, 2026, on the Kimi app, Kimi Code, and the Kimi API. Full open-weight model files are scheduled for public release by July 27, 2026.

4. How many parameters does KIMI K3 have?

KIMI K3 has approximately 2.8 trillion total parameters, organized across 896 experts in a Mixture-of-Experts architecture with 16 experts activated per token during inference.

5. What makes KIMI K3's architecture different?

KIMI K3 introduces Kimi Delta Attention (KDA), a hybrid linear-attention mechanism that combines standard attention with linear attention residuals, enabling efficient processing of one-million-token contexts without quadratic memory overhead.

6. What are the two KIMI K3 variants?

K3 Max is the general-purpose variant for chat, knowledge work, coding, and agentic tasks. K3 Swarm Max is optimized for large-scale parallel processing and high-throughput multi-agent workloads.

7. How does KIMI K3 benchmark against competitors?

KIMI K3 ranks fourth globally on overall frontier evaluation, trailing only Claude Fable 5 and GPT-5.6 Sol while surpassing Claude Opus 4.8. On coding benchmarks specifically, K3 surpasses Claude Fable 5.

8. What is KIMI K3's context window?

KIMI K3 supports a one-million-token context window, enabled by the KDA architecture that processes very long sequences efficiently.

9. What is KIMI K3's pricing?

Standard API pricing is $3 per million input tokens and $15 per million output tokens. Cached input tokens cost $0.30 per million, with Mooncake infrastructure designed to maintain above 90% cache hit rates for code-heavy workflows.

10. Is KIMI K3 open-weight?

Yes. Full model weights are scheduled for public release by July 27, 2026 under a Modified MIT license, making KIMI K3 the first open 3-trillion-class model available for download, self-hosting, and fine-tuning.

11. What is thinking mode in KIMI K3?

Thinking mode is KIMI K3's always-on reasoning capability. The model reasons through problems step-by-step before generating a final response, running by default across all interactions rather than requiring a separate reasoning variant.

12. Does KIMI K3 support images?

Yes. KIMI K3 includes native vision and multimodal capability, processing images through its core model architecture rather than an externally attached vision module.

13. What is K3 Swarm Max?

K3 Swarm Max is a KIMI K3 variant optimized for large-scale parallel processing, designed for environments where many simultaneous agentic tasks need to execute concurrently at high throughput.

14. How does KIMI K3 compare to DeepSeek V4 Pro?

KIMI K3 at 2.8 trillion parameters is approximately 75% larger than DeepSeek V4 Pro at 1.6 trillion and substantially outperforms it on benchmark evaluations, taking the title of the world's largest open-weight model.

15. Why does KIMI K3's coding benchmark result matter?

It is the first time an open-weight model has surpassed Claude Fable 5, currently the strongest publicly accessible proprietary model, on coding benchmarks. This demonstrates that the performance gap between open and closed models on coding has closed.

16. What can organizations do with KIMI K3's open weights?

Organizations can self-host KIMI K3 for data privacy and latency control, fine-tune it on proprietary data for domain specialization, use it for academic and research purposes, and build derivative models under the Modified MIT license.

17. What subscription tier is needed for KIMI K3 in the consumer app?

Access to K3 Max in the Kimi app begins at a ¥199 subscription tier. The app is available on iOS and Android.

18. What is the KIMI K3 API recharge promotion?

A promotion offering 10% to 30% bonus credits on API recharges is running from July 15 through August 11, 2026, with the bonus percentage varying by recharge amount.

19. What are KIMI K3's main limitations?

KIMI K3 trails Claude Fable 5 and GPT-5.6 Sol on general capability. Full weights are not yet publicly available. Video and screen sharing are not supported in the Kimi app at launch. Non-English, non-Chinese language performance should be independently validated.

20. How should teams evaluate KIMI K3 for their specific use case?

Run representative task-specific benchmarks on your actual workloads, particularly coding and long-context tasks where K3 is strongest. Test during the current API recharge promotion to reduce evaluation costs. Plan fine-tuning and self-hosting strategies around the July 27 open-weight release if domain adaptation or infrastructure control is a priority.