On July 16, 2026, Moonshot AI launched KIMI K3, the most powerful open-weight large language model ever released. At 2.8 trillion total parameters, it is the first open 3-trillion-class model in AI history, surpassing DeepSeek's previous record of 1.6 trillion parameters and arriving as a direct challenger to the most advanced proprietary systems from Anthropic and OpenAI. The launch was tipped a day early when a promotional page on Moonshot's own Kimi Open Platform leaked before the official announcement, immediately generating widespread attention across the global AI research and developer community.

KIMI K3 is not merely a parameter count story. It introduces a new architecture built on Kimi Delta Attention, a native vision and multimodal capability, an always-on reasoning mode called thinking mode, a one-million-token context window, and an open-weight release scheduled for July 27, 2026, under a Modified MIT license. On independent benchmarks, it ranks fourth among all frontier models globally, trailing only Claude Fable 5 and GPT-5.6 Sol on overall capability while edging past Claude Opus 4.8 and GPT-5.5.

For AI practitioners, researchers, and technology professionals who want to build genuine, expert-level understanding of frontier AI systems like KIMI K3, grounded in the technical and conceptual foundations that make these models work, a structured Certified Artificial Intelligence (AI) Expert certification provides exactly that depth, equipping professionals to evaluate, apply, and govern models across the rapidly evolving frontier AI landscape with real precision.

This guide covers the complete picture: what KIMI K3 is, how its architecture works, what it benchmarks, how it is priced, how it compares to the frontier, and why it matters for the global AI industry.

What Is KIMI K3?

KIMI K3 is Moonshot AI's flagship large language model and the third major generation in its Kimi model series, following the K1 family and the K2 series, which progressed through K2, K2.5, K2.6, and K2.7 Code. It was released on July 16, 2026, initially via the Kimi app, the Kimi Code platform, and the Kimi API, with full open weights scheduled to publish by July 27 under a Modified MIT license.

At launch, two variants are available:

K3 Max is the primary general-purpose variant, optimized for chat, knowledge tasks, complex coding, and multi-step agentic work. This is the model that runs in the Kimi app and API by default.

K3 Swarm Max is a second variant optimized for large-scale parallel processing workloads, designed for environments where many simultaneous agent tasks need to run efficiently at high throughput.

Moonshot's own positioning is precise: "While its overall performance still trails the most powerful proprietary models, Claude Fable 5 and GPT-5.6 Sol, Kimi K3 demonstrated frontier-level performance across our evaluation suite, consistently outperforming other tested models." That framing is notable for its honesty and for what it implies: KIMI K3 places Moonshot AI firmly in the global frontier tier, not as a competitor that almost gets there but as one that has demonstrably arrived.

What Makes KIMI K3 Different

Mixture-of-Experts at 2.8 Trillion Parameters

KIMI K3 uses a Mixture-of-Experts (MoE) architecture with approximately 2.8 trillion total parameters organized into 896 experts. At inference, 16 experts are activated per token. This means the model uses only a fraction of its total capacity for any given token, making frontier-scale reasoning economically viable at the parameter counts involved.

The MoE design also explains why KIMI K3 can be competitive on quality benchmarks while remaining practically deployable: activating 16 of 896 experts per token is far less computationally intensive per inference call than a dense model at the same parameter count.

Kimi Delta Attention (KDA)

The architectural centerpiece that distinguishes KIMI K3 from prior Kimi models and from most other frontier models is the Kimi Delta Attention mechanism. KDA is a hybrid linear-attention approach that combines standard attention with linear attention residuals.

In practical terms, this architecture change enables more efficient processing of very long contexts, specifically the one-million-token context window that KIMI K3 supports, without the quadratic memory cost that standard attention incurs at such lengths. This is a genuine architectural innovation, not an incremental scaling decision.

Native Vision and Multimodality

KIMI K3 ships with native multimodal capability, meaning it processes images directly rather than through a bolted-on vision module. Native visual understanding is integrated into the model's core reasoning pathway, which affects the quality of tasks that combine text reasoning with image interpretation: chart analysis, diagram reading, scientific figure interpretation, and document understanding.

Always-On Thinking Mode

Previous KIMI models offered a separate reasoning variant for tasks requiring step-by-step deliberation. KIMI K3 unifies this: thinking mode runs by default across all interactions. The model reasons through problems before responding rather than offering a direct completion and a reasoning completion as separate products. This design reflects the direction of the frontier: always-on reasoning is becoming the expected baseline, not an optional add-on.

KIMI K3 Benchmarks: Where It Stands

Moonshot AI published a full set of benchmark comparisons at launch, comparing KIMI K3 against Claude Fable 5, GPT-5.6 Sol, Claude Opus 4.8, and GPT-5.5. Independent analysis from Artificial Analysis and individual researchers corroborates the broad picture.

Overall Positioning

On overall frontier ranking, independent testing places KIMI K3 fourth among all models evaluated, behind Claude Fable 5 and GPT-5.6 Sol, and ahead of Claude Opus 4.8. In Moonshot's own evaluation suite, K3 consistently outperforms Claude Opus 4.8 maximum settings and GPT-5.5 high settings, while trailing Fable 5 and GPT-5.6 Sol.

Coding Performance

On coding benchmarks, KIMI K3's performance is its strongest category. The model's coding scores surpass Claude Fable 5, making coding the dimension on which K3 not only reaches the frontier but exceeds it. For developers and software engineering teams evaluating whether K3 is appropriate for production coding workflows, this result carries direct practical weight.

On SWE-bench Verified, K3 produces a high score that positions it competitively at the frontier for multi-file software engineering tasks. Exact published figures confirm it as among the top three models globally on this evaluation.

GPQA (Expert-Level Reasoning)

On GPQA, which tests expert-level scientific reasoning across biology, chemistry, and physics, KIMI K3 performs competitively with Opus 4.8 and surpasses GPT-5.5, while trailing Fable 5 and Sol. The science domain strength matters particularly for research-adjacent applications in life sciences, chemistry, and engineering where expert reasoning depth is required.

Long-Context Performance

Given KIMI K3's one-million-token context window, Moonshot tested long-context performance directly. The model maintains strong performance across very long documents and multi-document retrieval tasks, which is operationally significant for use cases involving large codebases, legal document review, and enterprise knowledge management.

KIMI K3 Pricing: API and Consumer Tiers

API Pricing

KIMI K3 is priced at $3 per million input tokens and $15 per million output tokens at standard rates. This pricing positions it directly in the same tier as Claude Sonnet 5 at standard rates and below Claude Opus 4.8 and GPT-5.6 Sol.

The most structurally significant pricing detail is the cached input cost: $0.30 per million cached input tokens, approximately one-tenth of the full input price. Moonshot's Mooncake serving infrastructure is designed to keep cache hit rates above 90% for coding workflows, which reduces the effective real-world input cost for code-heavy use cases by approximately four times compared to the nominal per-token rate.

This cached pricing structure, combined with coding performance that exceeds Claude Fable 5 on specific benchmarks, makes KIMI K3 an exceptionally strong cost-performance choice for software engineering teams running high-volume coding workflows.

Consumer Access

In the Kimi app, access to K3 Max is available starting at a ¥199 subscription tier. Access to KIMI K3 within the Kimi app includes iOS availability as of launch. A recharge promotion offering 10% to 30% bonus API credits is running from July 15 through August 11, 2026.

API Promotion

The launch recharge promotion applies to API credits recharged between July 15 and August 11, with bonus credit percentages varying by recharge amount. Teams evaluating KIMI K3 for production integration should consider taking advantage of the promotion window to run extensive benchmark testing at reduced effective cost before committing to full production volumes.

The Open-Weight Promise: What July 27 Means

Perhaps the most consequential dimension of KIMI K3's launch for the global AI research community is not its benchmark performance but its open-weight commitment. Full model weights are scheduled to publish by July 27, 2026 under a Modified MIT license, making KIMI K3 not just the largest open-weight model ever released but also one of the most commercially permissive frontier releases in AI history.

This means:

Researchers can download and study the weights directly, examining the MoE architecture, the KDA attention mechanism, and the training signature without relying on API access or inference-layer abstraction.

Organizations can fine-tune K3 on proprietary data, adapting the model to domain-specific tasks including specialized medical reasoning, financial analysis, legal document processing, and industrial applications.

Developers can self-host the model, removing the API dependency entirely for use cases where data privacy, latency control, or cost predictability require on-premises deployment.

The global research community gains access to a 3T-class architecture for the first time, enabling reproduction studies, architectural analysis, and second-order innovations that build on the K3 foundation.

The open-weight release also sets a direct precedent for how China's frontier AI labs are positioning themselves globally: not just as capability competitors to OpenAI and Anthropic but as open-source contributors to the global AI ecosystem, on terms significantly more permissive than typical proprietary releases.

How Moonshot AI Got Here: The $500 Million Series C

The engineering resources required to train a 2.8 trillion parameter model on a new architecture do not materialize without significant capital commitment. Moonshot AI closed a $500 million Series C round in January 2026 at a $4.3 billion valuation, explicitly earmarked for KIMI K3 development and compute infrastructure expansion.

That funding, combined with Moonshot's Mooncake serving infrastructure, which provides the high-throughput, cache-optimized inference environment required to serve a 2.8T MoE model at competitive API latencies and pricing, represents a substantial investment in both training and deployment infrastructure that forms the foundation for K3's commercial viability.

Moonshot was founded in March 2023 and released the original Kimi chatbot in October 2023, initially known for supporting up to 128,000 tokens of context, the first model to do so at that scale. Monthly active users exceeded 36 million by late 2024, providing both the user data and commercial revenue base that supported the K3 development cycle.

KIMI K3 vs. the Global Frontier: A Clear-Eyed Comparison

vs. Claude Fable 5

Claude Fable 5 is the current overall leader on most benchmark suites and retains a meaningful lead over KIMI K3 on general capability. However, on coding benchmarks specifically, KIMI K3 surpasses Fable 5. For organizations whose primary use case is software engineering, the capability gap narrows significantly and may reverse depending on the specific coding task evaluated.

vs. GPT-5.6 Sol

GPT-5.6 Sol leads KIMI K3 on the Terminal-Bench 2.1 agentic coding benchmark and on overall frontier evaluation. However, GPT-5.6 Sol is a proprietary, closed-weight model at higher pricing. KIMI K3 offers comparable performance in many dimensions at lower API cost with an open-weight release, a combination that no competing frontier model currently matches.

vs. Claude Opus 4.8

KIMI K3 outperforms Claude Opus 4.8 on the majority of benchmark categories evaluated. This is the most practically significant comparison for most enterprise teams: Opus 4.8 is the current mid-to-upper-tier proprietary default for many organizations, and K3's ability to exceed it as an open-weight model changes the cost-performance calculation for teams currently routing to Opus 4.8.

vs. DeepSeek V4 Pro (Previous Open-Weight Record)

KIMI K3 at 2.8 trillion parameters is 75% larger than DeepSeek V4 Pro's 1.6 trillion parameters, the previous largest open-weight model. On benchmark performance, K3 substantially outperforms V4 Pro across most evaluated categories.

What KIMI K3 Means for the Global AI Landscape

The Open-Weight Frontier Has Reached Parity on Coding

For the first time, an open-weight model has exceeded proprietary frontier models, specifically Claude Fable 5, on coding benchmarks. This is not a marginal result on a secondary benchmark. It is a direct signal that the performance gap between open and closed models, which defined the first generation of the frontier, is now definitively closing on specific task categories.

Chinese AI Labs Are Operating at Frontier Tier

KIMI K3's fourth-place global ranking and coding leadership confirm that Moonshot AI is not competing at a distance from the frontier. It is operating within the frontier tier. This matters for global AI strategy: the assumption that frontier AI capability is geographically concentrated is increasingly unsupported by the evidence.

Open-Weight Models Are Becoming Strategically Significant

A frontier-tier open-weight model changes the options available to every enterprise, university, research institution, and government that wants frontier AI capability without proprietary API dependency. The July 27 weight release will make KIMI K3 deployable on-premises by any organization with the compute infrastructure to run a 2.8T MoE model.

Use Cases Where KIMI K3 Excels

Enterprise Software Engineering: Given its coding benchmark leadership over Claude Fable 5 and strong SWE-bench performance, KIMI K3 is a directly applicable choice for large-scale software engineering workflows. The cached input pricing at $0.30 per million tokens, combined with Mooncake's 90%+ cache hit rate for code, makes it highly cost-effective for teams processing large codebases continuously.

Long-Document Analysis: The one-million-token context window makes KIMI K3 the most capable open-weight model available for tasks requiring the simultaneous processing of entire document libraries, lengthy codebases, multi-document research synthesis, or extended agentic sessions.

Research and Science: GPQA performance that surpasses GPT-5.5 positions K3 for legitimate use in expert-level scientific reasoning tasks, including literature synthesis, hypothesis generation, and technical problem-solving in biology, chemistry, and physics.

Multimodal Document Understanding: Native vision integration, not a bolted-on module, enables higher-quality processing of documents that combine visual and textual content: technical diagrams, scientific figures, financial tables, and engineering specifications.

Fine-Tuning and Domain Specialization: Once the July 27 open-weight release lands, organizations can fine-tune K3 on proprietary data for specialized domains. Healthcare, legal, financial, and industrial applications that require frontier capability on proprietary data without API exposure become directly feasible.

Building Expertise in the KIMI K3 Era

The arrival of KIMI K3 as a frontier-tier, open-weight model changes what it means to operate competently in AI-driven technology environments. For technology professionals seeking to understand how to evaluate, deploy, and build systems around frontier models at this level, the technical foundation matters as much as access to the tools themselves.

A comprehensive Tech Certification covering AI systems, model architectures, cloud infrastructure, and enterprise deployment patterns provides the structural literacy needed to evaluate KIMI K3's technical claims, understand MoE architecture trade-offs, assess the KDA attention mechanism's practical implications, and make informed decisions about when K3 is the right model for a specific production context.

Limitations Worth Knowing

Not yet at full parity on general tasks: While KIMI K3 leads on coding and is competitive across most benchmarks, Claude Fable 5 and GPT-5.6 Sol retain a meaningful general capability lead. Teams should run task-specific evaluations before defaulting to K3 for use cases outside its strongest benchmark categories.

Full weights not yet available: The July 27 weight release is scheduled but not yet complete. Organizations planning to self-host or fine-tune K3 need to plan around that date rather than treating the weights as available today.

No video or screen sharing at launch: Unlike some competing platforms, the Kimi app does not support real-time video or screen interaction with K3 at the July 16 launch. This limits the voice and multimodal interaction capabilities compared to platforms like Gemini Live.

Language coverage limitations: K3's training data and benchmark evaluation are strongest in Chinese and English. Teams requiring frontier-tier performance in other languages should test performance in their specific language before committing to production deployment.

Building Business Fluency Around KIMI K3

Understanding the technical dimensions of KIMI K3 is one dimension of professional readiness. For practitioners who also need to communicate K3's strategic implications to leadership, evaluate its implications for AI procurement strategy, or position AI investments to non-technical stakeholders, the ability to frame these developments in business language is equally important. A Marketing Certification develops the strategic communication, business positioning, and stakeholder management skills that allow technology professionals to translate frontier model developments like KIMI K3 into clear, credible organizational guidance alongside a Tech Certification and a Certified Artificial Intelligence (AI) Expert credential.

Conclusion

KIMI K3 is the most important open-weight model release in the history of AI. At 2.8 trillion parameters, it is the largest open-weight model ever built. On coding benchmarks, it surpasses Claude Fable 5 while remaining priced below Opus 4.8 and GPT-5.6 Sol. Its open-weight release on July 27 under a Modified MIT license will make frontier-tier AI capability available for self-hosting and fine-tuning to any organization with the infrastructure to deploy it.

For AI practitioners, developers, researchers, and enterprise technology teams, KIMI K3 represents both an immediately usable frontier model and a structural shift in who has access to frontier-grade AI capability. Building the expertise to evaluate and apply it responsibly, through a Certified Artificial Intelligence (AI) Expert certification, a Tech Certification grounding in AI systems and infrastructure, and a Marketing Certification for strategic communication, positions professionals and organizations to operate confidently at the new frontier that KIMI K3 has established.

FAQs

1. What is KIMI K3?

KIMI K3 is Moonshot AI's flagship large language model, launched July 16, 2026. It is the world's largest open-weight AI model at approximately 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 made KIMI K3?

KIMI K3 was developed by Moonshot AI, a Chinese AI startup founded in March 2023, 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 was officially released on July 16, 2026, initially on the Kimi Code platform and the Kimi app. Full open model weights are scheduled for public release by July 27, 2026.

4. What are KIMI K3's two variants?

K3 Max is the primary general-purpose variant for chat, knowledge work, coding, and agentic tasks. K3 Swarm Max is a second variant optimized for large-scale parallel processing at high throughput.

5. How large is KIMI K3?

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.

6. What is Kimi Delta Attention (KDA)?

KDA is KIMI K3's custom hybrid linear-attention architecture that combines standard attention with linear attention residuals. It enables efficient processing of one-million-token context windows without the quadratic memory cost of standard attention at that length.

7. How does KIMI K3 perform on benchmarks?

On independent evaluations, KIMI K3 ranks fourth globally among all frontier models, 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, making it capable of processing entire codebases, large document libraries, and extended agentic sessions within a single context.

9. What is KIMI K3's API pricing?

KIMI K3 is priced at $3 per million input tokens and $15 per million output tokens. Cached input tokens cost $0.30 per million, with Mooncake infrastructure maintaining above 90% cache hit rates for code-heavy workflows.

10. Is KIMI K3 open-source?

KIMI K3 is an open-weight model. Full model weights are scheduled for public release by July 27, 2026 under a Modified MIT license, making it 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, where the model reasons through problems step-by-step before generating a final response. Unlike previous Kimi models that offered reasoning as a separate variant, K3 applies thinking mode by default across all interactions.

12. Does KIMI K3 support images and vision?

Yes. KIMI K3 includes native vision and multimodal capability, processing images directly within the model's core architecture rather than through a separately attached vision module.

13. How does KIMI K3 compare to Claude Opus 4.8?

KIMI K3 outperforms Claude Opus 4.8 on the majority of benchmark categories evaluated. For teams currently routing production workloads to Opus 4.8, K3 offers a compelling performance improvement at comparable or lower pricing, with the additional advantage of open weights.

14. How does KIMI K3 compare to Claude Fable 5?

Overall, Claude Fable 5 retains a meaningful general capability lead. However, on coding benchmarks specifically, KIMI K3 surpasses Fable 5, making it the stronger choice for software engineering-focused deployments.

15. What is K3 Swarm Max?

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

16. What subscription tier is needed to access KIMI K3 in the app?

Access to K3 Max in the Kimi app begins at a ¥199 subscription tier. The app is available on iOS as of the July 16, 2026 launch.

17. What is the KIMI K3 API recharge promotion?

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

18. What makes KIMI K3 historically significant?

It is simultaneously the largest open-weight model ever released (2.8 trillion parameters), the first open 3-trillion-class model, and the first open-weight model to surpass a leading proprietary frontier model (Claude Fable 5) on coding benchmarks.

19. What are KIMI K3's current limitations?

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

20. How should enterprise teams evaluate KIMI K3 for production use?

Run representative task-specific benchmarks against your actual use cases, particularly coding and long-context tasks where K3 is strongest. Test the API with the current recharge promotion to assess quality and latency before committing to production volumes. Plan fine-tuning strategies around the July 27 open-weight release if domain adaptation is a priority.