Key Takeaways
- Kimi K3 is a 2.8-trillion-parameter open-weight AI model from Moonshot AI, released July 16, 2026, and the largest open-source model in existence. It uses a Mixture-of-Experts architecture with 896 experts, activating 16 per inference, paired with a 1-million-token context window, native vision capabilities, and two architectural innovations, Kimi Delta Attention and Attention Residuals, that deliver approximately 2.5x better scaling efficiency compared to its predecessor Kimi K2.
- The open weights were published on July 27, 2026, on Hugging Face (moonshotai/Kimi-K3) as 96 shards totalling approximately 1.56 TB. The release includes the full 2.8T MoE model, optimized attention kernels, MoE communication libraries, and deployment infrastructure components for self-hosting. The weights are distributed under a custom Kimi K3 License, not MIT, with a commercial revenue threshold that requires a separate agreement with Moonshot AI for entities and affiliates whose combined revenue exceeds $20 million over any consecutive 12 months.
- On coding and agentic benchmarks, Kimi K3 places first on Program Bench, SWE Marathon, BrowseComp, Automation Bench, Frontend Code Arena, and OmniDocBench. On Frontend Code Arena, a blind developer voting evaluation, K3 scores 1,679 points, placing it first ahead of Claude Fable 5. On SWE-bench Verified, it scores 76.8%. It consistently outperforms Claude Opus 4.8 and GPT-5.5 across the full evaluation suite.
- Kimi K3 trails Claude Fable 5 and GPT-5.6 Sol on overall capability rankings. Claude Fable 5 leads by approximately 10 points on HLE-Full reasoning, and GPT-5.6 leads by 5.5 points on DeepSWE. The Epoch Capabilities Index composite positions K3 behind both of those models in overall ranking. Its lead over Western models concentrates in frontend and agentic coding tasks rather than general reasoning.
- API pricing is $3 per million input tokens and $15 per million output tokens, with cached input priced at $0.30 per million tokens. This makes K3 approximately 60% cheaper than Claude Opus 5 ($5/$25) and 70% cheaper than Claude Fable 5 ($10/$50) on a per-token basis, while delivering competitive coding performance against both. The flat pricing applies across the full 1M-token context window.
- Self-hosting from open weights eliminates data sovereignty concerns that the API cannot address. For enterprises with data residency requirements, regulated-industry compliance constraints, or security policies against sending data to a Chinese-owned API endpoint, the open weights make Kimi K3 a deployable frontier-class model on their own infrastructure.
- Kimi K3 is the first open-weight model to reach what the industry calls the 3T-class, demonstrating that the performance gap between open-source and proprietary frontier models is narrowing to a range of 10 percentage points or less on core benchmarks, rather than the wide gap that existed in 2024 and early 2025. Its release is significant not only for its own performance but for what it signals about the trajectory of open AI development.
Since early 2025, the running assumption in AI infrastructure planning has been that frontier-level performance is available only from proprietary models with closed weights. Kimi K3 challenges that assumption directly. At 2.8 trillion parameters: it is the largest open-weight model ever released. On several coding and agentic benchmarks, it places first ahead of Claude Fable 5. And its full weights are available on Hugging Face for any team to download and run on their own hardware.
This review covers the architecture, benchmarks, pricing, licensing, and practical considerations for anyone evaluating Kimi K3 for production use.
What Is Kimi K3?
Kimi K3 is the flagship model from Moonshot AI, a Chinese AI laboratory. It was released for API access on July 16, 2026, with open weights published on July 27, 2026. The model represents a significant scale increase from its predecessor: at 2.8 trillion total parameters and 896 MoE experts: it is the first openly available model to reach the 3T-class scale, a threshold previously held exclusively by undisclosed large proprietary models from major Western AI laboratories.
Moonshot AI positioned K3 as a model designed to compete with the best Western proprietary systems on tasks that matter most for enterprise AI deployment: software engineering, agentic reasoning, long-document analysis, and frontend development. The model includes native vision capabilities, accepting images (and reportedly video) as input natively without a separate multimodal pipeline.
The release of the full open weights is the element that most distinguishes Kimi K3 from prior open-weight releases. At 1.56 TB distributed across 96 shards, it requires substantial compute infrastructure to run, but for organizations with the hardware, it provides frontier-class AI performance on fully self-hosted infrastructure, with complete control over the inference environment, data handling, and customization.
Architecture and Technical Design
Mixture-of-Experts at Scale
Kimi K3 uses a Mixture-of-Experts architecture with 896 experts and 16 active experts per inference. This means the full 2.8 trillion parameters are not all active during any single inference pass. Only 16 expert networks process each token. The total parameter count determines the model’s potential capacity and specialization; the active expert count determines the compute cost per request. This architecture allows K3 to encode significantly more knowledge and specialization than a dense model of the same inference cost would permit.
The MoE implementation uses a Stable LatentMoE framework, which Moonshot designed to address instability issues that occur at scale in standard MoE training. At 896 experts, standard MoE routing can produce expert collapse (where most routing goes to a small subset of experts) or training instability. The Stable LatentMoE framework maintains expert utilization and training stability across the full 896-expert pool.
Kimi Delta Attention and Attention Residuals
Two architectural innovations distinguish K3’s transformer design from standard attention implementations. Kimi Delta Attention (KDA) modifies how attention operates across sequence length, designed to improve information flow in long-context scenarios without the quadratic compute cost that standard attention requires at 1M-token lengths. Attention Residuals (AttnRes) modify how attention information is propagated across model depth, improving gradient flow during training and representation quality at inference.
Together, Moonshot reports these two changes deliver approximately 2.5x better scaling efficiency compared to Kimi K2, meaning K3 achieves higher benchmark performance per parameter and per FLOP of training compute than the previous generation. The practical effect is that K3 punches above its effective compute weight: its inference cost per token (based on 16 active experts) is substantially lower than the 2.8T headline parameter count would suggest.
Context and Multimodal
The context window is 1 million tokens, flat across all pricing tiers. Native vision capability allows image input without a separate model route. Moonshot has indicated video input support is in development, with image support confirmed at launch. The 1M-token context makes K3 suitable for the same long-document analysis use cases as Claude Opus 5 and Gemini 2.5 models.
Benchmark Performance
Coding and Agentic Tasks
Kimi K3’s benchmark lead over Western models concentrates in software engineering and agentic coding evaluations. It places first on six major benchmarks: Program Bench, SWE Marathon, BrowseComp, Automation Bench, Frontend Code Arena, and OmniDocBench. The Frontend Code Arena result is the most notable: in blind developer voting where coders evaluate AI-generated frontend code quality without knowing which model produced it, K3 scores 1,679 Arena points, placing it first ahead of Claude Fable 5. This evaluation type is harder to game than automated benchmarks and reflects actual developer preference on the specific task type.
On SWE-bench Verified, K3 scores 76.8%. This is lower than Claude Opus 5’s 96.0% on the same benchmark, a substantial gap that reflects SWE-bench’s broader task coverage versus K3’s more concentrated frontend and agentic coding strengths. The difference between 76.8% on SWE-bench Verified and first place on Frontend Code Arena suggests K3’s coding performance is strongest on frontend and UI-focused engineering tasks rather than the full range of backend and systems-level tasks that SWE-bench includes.
Reasoning and General Capability
On HLE-Full reasoning, Claude Fable 5 leads K3 by approximately 10 percentage points. On DeepSWE (a harder software engineering evaluation), GPT-5.6 Sol leads by 5.5 points. On overall composite capability rankings (Epoch Capabilities Index), K3 places behind both Fable 5 and GPT-5.6.
The pattern across benchmarks is consistent: K3 is competitive with or ahead of the best proprietary models on frontend coding, agentic task completion, and document analysis. It trails on tasks requiring the deepest reasoning, mathematical abstraction, and full-spectrum software engineering. This concentration of strength is a practical consideration for teams evaluating it: K3 is not a general-purpose replacement for Fable 5 or GPT-5.6 across all task types, but it is a competitive or superior option for the specific workloads where it leads.
Pricing and Access
API Pricing
Kimi K3 API pricing is $3 per million input tokens and $15 per million output tokens, with cached input at $0.30 per million tokens. This pricing is flat across the full 1M-token context window, there is no surcharge for using the extended context. Compared to Claude Opus 5 at $5/$25 per million tokens, K3 costs 40% less on input and 40% less on output. Compared to Claude Fable 5 at $10/$50 per million tokens, K3 costs 70% less on both sides.
For cost-sensitive production deployments where the workload concentrates on tasks where K3 performs competitively, the pricing gap is significant. A team running 100 million tokens monthly saves $2,000/month versus Opus 5 on input alone, or $700,000 annually versus Fable 5 on input and output combined at equivalent volume.
Open Weights Self-Hosting
The full model weights are available on Hugging Face at moonshotai/Kimi-K3 as of July 27, 2026. The download is 96 shards totalling approximately 1.56 TB, including inference infrastructure components, optimized attention kernels, and MoE communication libraries. Running K3 at full scale requires substantial GPU cluster resources: the 2.8T parameter model at 16-active-expert inference is not deployable on consumer hardware but is feasible on enterprise GPU infrastructure.
For teams with the infrastructure, self-hosting eliminates per-token API costs entirely. The cost of running a self-hosted K3 deployment becomes a function of GPU compute and infrastructure, rather than token volume. For high-volume production deployments that justify the infrastructure investment, self-hosting can deliver frontier-class performance at significantly lower marginal cost than any API offering.
Licensing Considerations
The open weights are released under a custom Kimi K3 License, not a standard open-source license like MIT or Apache 2.0. For most research and commercial deployments, the license is permissive. The key commercial threshold: organizations whose combined revenue (including affiliates) exceeds $20 million over any consecutive 12-month period and who provide inference or fine-tuning access to third parties with control over inputs and parameters must enter a separate commercial agreement with Moonshot AI. This threshold does not affect internal deployment or research use. Organizations evaluating K3 for production should review the license terms for their specific deployment structure before committing to a self-hosted production architecture.
Data Sovereignty and the China Question
Kimi K3 is a product of a Chinese AI laboratory. This creates a set of considerations that are separate from its technical performance. For organizations in regulated industries, defense contracting, government, or with data residency requirements, the question of whether to send data to a Chinese-owned API endpoint is a compliance and policy question independent of model quality.
The open weights address this concern directly for organizations with the infrastructure to act on it. Running Kimi K3 from self-hosted open weights sends no data to Moonshot AI’s systems. Data remains entirely on the organization’s infrastructure, making K3 potentially viable in environments where API-based access would be prohibited. This is a meaningful distinction from proprietary models, where no comparable self-hosting option exists regardless of data sensitivity.
For teams without data sovereignty constraints using the API, the consideration is smaller. Moonshot’s API infrastructure is operated commercially and is subject to the same category of trust considerations as any third-party AI API. Teams should apply their standard vendor due diligence processes rather than treating origin country as the sole relevant factor.
Who Should Use Kimi K3
Kimi K3 is the strongest current option for teams whose primary AI workload is frontend development and UI engineering. Its first-place position on Frontend Code Arena in blind developer evaluation reflects genuine performance quality on this task type. For frontend-focused teams or companies building frontend code generation tools, K3 outperforms Claude Fable 5 on the specific task that matters most to them while costing 70% less per token.
Research teams and organizations that require open-weight model access for compliance, customization, fine-tuning, or self-hosting have the clearest use case for K3’s open weight release. No proprietary model at competitive frontier performance levels is available on open weights. K3 fills that gap for organizations where proprietary API access is not viable.
Cost-sensitive production deployments running agentic coding, document analysis, or BrowseComp-style web navigation tasks can run K3 at 40% to 70% lower cost than Anthropic or OpenAI flagship models while maintaining competitive performance on those task types. Teams already using GPT-5.5 or Claude Opus 4.8 who are primarily running coding or agentic workflows should benchmark K3 against their current model before renewing at existing price points.
Teams whose workload centers on deep mathematical reasoning, general-purpose question answering, or the hardest coding tasks should benchmark carefully before adopting K3. Its 10-point gap behind Fable 5 on HLE-Full reasoning and lower SWE-bench Verified score relative to Claude Opus 5 mean that for these task types, Western frontier models maintain a meaningful performance advantage that the cost savings may not fully offset.
Verdict
Kimi K3 is the most significant open-weight model release since Llama 3, and the first to credibly compete with frontier proprietary models on specific high-value task types rather than merely reducing the gap. Its first-place ranking on Frontend Code Arena ahead of Claude Fable 5, combined with open weights and API pricing 40 to 70% below Western competitors, makes it a serious production option for the workloads where it leads.
The 2.8 trillion parameters and the open-weight release together represent something the AI industry has not seen before: a model at the technical frontier, from outside the US AI cluster, available for anyone to download and run. Whether that is a reason to adopt it depends on the specific task, the infrastructure, the compliance environment, and the risk tolerance of each organization. But it is a reason to evaluate it seriously rather than dismiss it based on its origin.
Frequently Asked Questions
What is Kimi K3?
Kimi K3 is a 2.8-trillion-parameter open-weight AI model developed by Moonshot AI, a Chinese artificial intelligence laboratory. It is the largest open-source AI model ever released, using a Mixture-of-Experts architecture with 896 experts and 16 active per inference. It was released for API access on July 16, 2026, with open weights published on Hugging Face on July 27, 2026. The model includes native vision capabilities and a 1-million-token context window.
How many parameters does Kimi K3 have?
Kimi K3 has 2.8 trillion total parameters, making it the first openly available model to reach the 3T-class scale. Due to its Mixture-of-Experts architecture, only 16 of its 896 experts are active during any single inference pass, which means the effective compute cost per token is substantially lower than the 2.8T total parameter count would suggest. The active parameter count during inference is the cost-relevant figure; the total parameter count reflects the model’s capacity and specialization depth.
Is Kimi K3 open source?
Kimi K3 weights are open and available on Hugging Face at moonshotai/Kimi-K3. However: they are distributed under a custom Kimi K3 License rather than a standard open-source license such as MIT or Apache 2.0. The license permits research and most commercial use. Organizations whose combined revenue with affiliates exceeds $20 million over 12 months and who provide inference or fine-tuning access to third parties with parameter control must enter a separate agreement with Moonshot AI. The weights are not fully open-source in the traditional sense but are freely available for download and self-hosting for most use cases.
How does Kimi K3 compare to Claude Fable 5?
Kimi K3 outperforms Claude Fable 5 on Frontend Code Arena (1,679 points vs Fable 5’s lower score in blind developer voting), and places ahead on Program Bench, SWE Marathon, BrowseComp, Automation Bench, and OmniDocBench. Claude Fable 5 leads on HLE-Full reasoning by approximately 10 percentage points and on overall composite capability indexes. On SWE-bench Verified, Fable 5 scores 80.3% and K3 scores 76.8%. The pattern is that K3 leads on frontend coding and agentic task completion; Fable 5 leads on deep reasoning and general capability. K3 costs 70% less per token via API.
What is Kimi K3’s pricing?
Kimi K3’s API pricing is $3 per million input tokens and $15 per million output tokens, with cached input at $0.30 per million tokens. This pricing is flat across the full 1-million-token context window. Compared to Claude Opus 5 at $5/$25 per million tokens, K3 is approximately 40% cheaper on both input and output. Compared to Claude Fable 5 at $10/$50 per million tokens, K3 is approximately 70% cheaper. Self-hosting from open weights eliminates per-token costs entirely, with cost determined by compute infrastructure instead.
What architecture does Kimi K3 use?
Kimi K3 uses a Mixture-of-Experts architecture with 896 total experts and 16 active per inference. It implements two architectural innovations developed by Moonshot AI: Kimi Delta Attention (KDA), which improves information flow across long sequences, and Attention Residuals (AttnRes), which improve gradient flow across model depth. These changes are implemented alongside a Stable LatentMoE framework that maintains expert utilization stability across the full 896-expert pool. Moonshot reports that K3’s architecture delivers approximately 2.5x better scaling efficiency compared to Kimi K2.
Can Kimi K3 handle images?
Yes. Kimi K3 includes native vision capabilities, accepting image inputs natively without a separate multimodal pipeline. This is available at the same API pricing across the 1-million-token context window. Video input support is reported to be in development. The native vision capability makes K3 applicable for tasks requiring visual understanding, document analysis from images, UI screenshot interpretation, and other vision-language tasks without requiring a separate model call or routing.
Where does Kimi K3 underperform compared to Western frontier models?
Kimi K3 trails Claude Fable 5 by approximately 10 percentage points on HLE-Full reasoning evaluation, and trails GPT-5.6 Sol by 5.5 points on DeepSWE (harder software engineering tasks). On overall composite capability indexes (Epoch Capabilities Index), it ranks below both Fable 5 and GPT-5.6. Its SWE-bench Verified score of 76.8% is below Claude Opus 5’s 96.0%. For tasks requiring deep mathematical reasoning, hard software engineering challenges beyond frontend focus, or general-purpose question answering at the highest difficulty level, Western frontier models maintain a performance advantage.
Is it safe to use Kimi K3 for sensitive business data?
The API safety question depends on an organization’s specific data policies and risk tolerance. Kimi K3 is a product of a Chinese AI laboratory, which may be a compliance or policy concern for organizations in regulated industries, government, or defense. For these organizations, the open weights provide a path to self-hosted deployment on internal infrastructure with no data sent to Moonshot AI’s systems. Teams without data sovereignty constraints should apply their standard vendor due diligence processes to API-based use of K3, as they would with any third-party AI API provider.
