Kimi K3: Why Moonshot AI’s 2.8-Trillion-Parameter Model Matters
Chinese artificial intelligence company Moonshot AI has released Kimi K3, a native multimodal model built for coding, research and long-running agent tasks. Independent evaluations place it near the global frontier, but its downloadable weights and full technical report remain unpublished, while unexpectedly high demand has already forced Moonshot to restrict new subscriptions.
Moonshot AI announced Kimi K3 on 16 July 2026 as a 2.8-trillion-parameter model with a one-million-token context window and native visual capabilities. The Beijing company says K3 is designed to complete complex coding, research and knowledge-work projects with limited human supervision. Independent evaluations indicate that it competes with leading models from Anthropic and OpenAI, although it does not clearly outperform them across every test. Its model weights are scheduled for release on 27 July and were not publicly downloadable at the time of publication.
What this article covers
This article explains how Kimi K3 works, what its 2.8-trillion-parameter scale means, how its independent benchmark results compare with Moonshot’s marketing claims, why its promised open-weight release matters, what caused its launch-week capacity shortage and whether K3 should be considered a continuously learning AI system.
Key takeaways
Kimi K3 contains 2.8 trillion total parameters, but its mixture-of-experts architecture activates only a small selection of specialised components for each piece of input.
The model supports a one-million-token context window and is designed for coding, research, visual reasoning and agentic workflows.
Independent evaluations place K3 among the leading current models, but its position varies by benchmark and several results remain preliminary.
Moonshot plans to release the complete weights on 27 July 2026. Until then, K3 is available as a hosted product and API rather than a downloadable open-weight model.
The final licence, full technical report, model card and safety evaluations have not yet been published.
Moonshot temporarily paused new consumer subscriptions after demand approached the limits of its computing clusters.
K3 does not appear to update its model weights continuously after deployment. Long context, external memory and repeated agent actions are not the same as continual learning.
What has Moonshot AI released?
Kimi K3 is the latest flagship model from Moonshot AI, the Beijing artificial intelligence company founded in 2023 by researcher Yang Zhilin.
The model is available through the Kimi website, the Kimi Work desktop application, the Kimi Code terminal agent and Moonshot’s developer API. Moonshot describes it as its most capable system so far, with particular strengths in software engineering, knowledge work, visual reasoning and tasks requiring many consecutive actions.
The launch moves Kimi beyond the conventional chatbot format. Moonshot’s demonstrations show the model navigating large code repositories, using terminal tools, producing editable presentations, creating interactive research reports and coordinating multiple subagents.
One demonstration involved constructing a small compiler. Another involved using open-source electronic-design software to develop and simulate a chip for a compact model. Moonshot also says K3 performed a lengthy computational astrophysics project involving academic papers, thousands of lines of code and hundreds of equations of state.
These demonstrations show the kind of work Moonshot wants K3 to perform. They do not independently prove that every user will obtain similar results. Long autonomous projects are sensitive to tool configuration, prompting, available compute, verification procedures and the amount of human intervention permitted.
What does 2.8 trillion parameters mean?
Parameters are the numerical values adjusted while a neural network is trained. They help determine how the system responds to language, images and other inputs.
A 2.8-trillion-parameter figure is unusually large, but it does not mean every one of those parameters is used for every word the model generates.
K3 is a mixture-of-experts model. Instead of sending every token through the entire network, a routing system selects a smaller group of specialised components, known as experts. Moonshot says K3 has 896 routed experts and activates 16 of them at a time, alongside shared components.
This design allows a model to contain a very large total capacity without paying the full computational cost of activating all 2.8 trillion parameters for every step. It is similar in principle to consulting a few relevant specialists rather than asking an entire organisation to work on every question.
The total parameter count still matters. A larger pool of experts can potentially support greater specialisation and capacity. It also increases storage, communication and deployment requirements.
Parameter count is not a reliable league table on its own. Architecture, training data, optimisation, post-training, inference strategy, tool access and evaluation conditions can matter as much as raw scale. A smaller model can outperform a larger one if it uses its parameters more effectively.
How is Kimi K3 different from Kimi K2?
Kimi K2, released in 2025, contained approximately one trillion parameters and activated 32 billion during inference. It established Moonshot as a serious developer of agent-focused open models, particularly for coding and tool use.
K3 increases the total model size to 2.8 trillion and introduces several architectural changes.
Kimi Delta Attention
Attention is the mechanism that allows a model to identify relationships between different parts of an input. Traditional attention becomes increasingly expensive as prompts grow longer.
Moonshot says Kimi Delta Attention provides a more efficient basis for processing long sequences. The company has not yet published enough technical detail for independent researchers to assess all of its claimed advantages.
Attention Residuals
Conventional transformer models pass information forward through layers using residual connections. Moonshot says its Attention Residuals system can selectively retrieve representations from different depths instead of accumulating them in a uniform way.
In simpler terms, the model is designed to choose which earlier internal representations remain useful as computation progresses.
Stable LatentMoE
K3 uses a highly sparse mixture-of-experts system with 896 routed experts. Moonshot says its Stable LatentMoE and balancing methods are intended to prevent a small number of experts from receiving too much work while others remain underused.
Quantisation-aware training
Moonshot says it incorporated low-precision numerical formats during post-training rather than attempting to compress the model only after development. This can make inference more efficient and improve compatibility with different classes of AI hardware.
The company reports that the complete collection of changes produces an approximate 2.5-fold improvement in scaling efficiency compared with Kimi K2. That remains a company result pending the full technical report and independent analysis.
What does the one-million-token context window allow?
A context window is the quantity of information a model can consider during one interaction.
K3’s one-million-token window can theoretically accommodate large software repositories, lengthy collections of documents or extended tool-use histories. It is substantially larger than the 256,000-token context supported by several earlier Kimi models.
This is useful for long projects because the model can retain more instructions, code, source material and previous actions within its active working context.
It does not guarantee perfect recall.
Models can overlook information, give disproportionate attention to particular sections or become less reliable as context expands. A one-million-token capacity describes the maximum permitted input, not flawless comprehension of every detail within it.
It is also not the same as permanent memory. Once material falls outside the context or a session ends, it does not automatically become part of the model’s underlying knowledge.
How strong is Kimi K3 in independent testing?
The early independent evidence is substantial enough to take K3 seriously, but not uniform enough to declare it the best model in the world.
Artificial Analysis gave K3 a score of 57 on its Intelligence Index, placing it third at the time of testing. The evaluator found its overall performance comparable with Claude Opus 4.8 and GPT-5.5, while GPT-5.6 Sol and Claude Fable 5 remained ahead.
Artificial Analysis also found that K3 generated approximately 130 million tokens during its evaluation, compared with an average of 63 million among comparable systems. This suggests the model may reach strong results partly through longer reasoning and more extensive output, which can affect cost and latency.
Vals AI gave K3 a score of 74.70 per cent on its overall index and placed it second among 38 systems. Its tests cover several professional and software-development tasks rather than relying on one general question-answering benchmark.
LM Arena placed K3 first on its web-development leaderboard, ahead of Claude Fable 5, following public comparisons of generated interfaces. The result was labelled preliminary and was based on 1,757 votes when reviewed. Rankings based on user preference can change as more votes are recorded.
Together, these results support three conclusions:
K3 belongs within the current frontier group.
Its strongest performance appears concentrated in coding, web development and agentic work.
It has not established universal superiority over Anthropic, OpenAI or every Chinese competitor.
Why Moonshot’s own benchmark table needs context
Moonshot publishes an extensive benchmark comparison covering coding, terminal work, office documents, web research and multimodal tasks.
The company is unusually open about several limitations in its methodology. Different models were sometimes tested through different agent harnesses. K3 frequently used Kimi Code, while OpenAI models used Codex and Anthropic models used Claude Code or other tools.
These differences matter because a coding agent is more than its underlying model. Performance can depend on:
The system prompt
File-editing tools
Context management
Error recovery
Terminal access
Search
Planning logic
Retry policies
The amount of allowed reasoning
The cost limit
Moonshot also reports that Claude Fable 5 encountered fallback behaviour on 35 per cent of tasks in one software-engineering evaluation, potentially reducing its measured performance. In another comparison, Moonshot tested models on hardware or modified tasks that differed from an official benchmark configuration.
These qualifications do not invalidate the results. They mean the headline scores should not be treated as perfectly controlled model-to-model comparisons.
The strongest evidence comes where independent evaluators reproduce K3’s performance using their own methods.
Is Kimi K3 open source?
Not yet in the broad sense, and not yet open-weight in the practical sense.
Moonshot says it will release K3’s full model weights by 27 July 2026. As of 20 July, users could access the model through Moonshot’s hosted products and API, but could not download the K3 weights from its public model repository. Independent model directories therefore continued to classify the available version as proprietary.
“Open-weight” generally means that the numerical parameters produced during training can be downloaded and run by others.
“Open source” can imply something broader, potentially including:
Training code
Inference code
Model architecture
Training data
Data-cleaning methods
Evaluation tools
A licence permitting modification and redistribution
Publishing weights alone does not necessarily provide all of these.
The final assessment will depend on what Moonshot releases on 27 July and which licence accompanies it. Restrictions on commercial use, competitive training or redistribution could significantly affect how open the model is in practice.
Could developers run K3 themselves?
Technically, a weight release could allow external deployment. Practically, the model’s scale creates a formidable barrier.
Moonshot recommends supernode configurations containing at least 64 accelerators. The model uses an extremely large number of parameters and depends on high-bandwidth communication between devices.
This is not a model that an ordinary developer will download onto a laptop or a single workstation.
Self-hosting is more likely to be realistic for:
Hyperscale cloud providers
Specialist AI infrastructure companies
National research laboratories
Large technology businesses
Well-funded universities
Governments and sovereign compute programmes
Most developers will continue using K3 through an API or a third-party inference provider even after the weights become available.
This limits one of the commonly assumed benefits of open-weight AI. A model can be legally downloadable while remaining economically inaccessible to most of the people entitled to download it.
How much does Kimi K3 cost?
Moonshot’s official API charges:
$0.30 per million cached input tokens
$3 per million uncached input tokens
$15 per million output tokens
The lower cached-input rate applies where previously processed context can be reused. Moonshot says its architecture achieves cache-hit rates above 90 per cent in some coding workloads.
K3 is cheaper than some leading proprietary systems in blended independent comparisons, but it is not automatically the lowest-cost option.
Artificial Analysis estimated a blended K3 price of approximately $2.31 per million tokens under its standard usage assumptions. Its comparison found K3 cheaper than GPT-5.6 Sol but more expensive than Z.ai’s GLM-5.2.
Output behaviour also matters. A model that generates more reasoning tokens may cost more to complete a task even when its published per-token price is lower.
Businesses should therefore compare complete task cost rather than input prices alone.
Why did Moonshot pause subscriptions?
On 20 July, Moonshot said it had temporarily stopped accepting new consumer subscriptions after demand exceeded its forecasts and approached the capacity limits of its computing clusters.
The company prioritised existing paid users and said it would reopen subscription places in batches as more capacity became available. It also announced plans to separate its general Kimi membership from a coding-focused plan, allowing compute to be allocated more precisely.
The interruption is evidence of strong interest. It is not, by itself, proof that K3 is technically superior. New AI releases frequently produce sudden traffic spikes, and capacity problems can reflect underprovisioning as much as exceptional product quality.
It nevertheless reveals an important constraint.
Long-horizon coding and agent tasks can require repeated model calls, large contexts and lengthy outputs. These workflows consume considerably more inference capacity than a brief chatbot answer. K3’s scale amplifies that requirement.
The release therefore illustrates the difference between training a frontier model and operating it reliably for millions of users.
What does Kimi K3 mean for China’s AI industry?
K3 adds to a series of increasingly competitive models from Chinese developers including DeepSeek, Alibaba, Z.ai and MiniMax.
Nuvastra previously examined how Chinese AI companies were narrowing the performance gap while competing through open-weight releases and lower prices. K3 strengthens that trend by moving a Chinese-developed system close to the top of several international evaluations.
This does not mean China has definitively overtaken the United States in artificial intelligence.
The leading position changes by benchmark. American companies continue to control substantial computing infrastructure, semiconductor design, cloud platforms and widely used commercial AI products. Several American developers also disclose less information about parameter counts, making direct scale comparisons difficult.
K3 does demonstrate that nationality is becoming a poor shortcut for estimating model quality.
Developers choosing an AI system are increasingly comparing specific performance, price, openness, latency, data governance and infrastructure requirements rather than assuming that the strongest model must come from one country.
Does Kimi K3 learn continuously after deployment?
Nothing in Moonshot’s published material indicates that K3 continuously updates its internal model weights in response to each new user, task or failure.
K3 can operate for long periods, retain a large working context, call tools, use stored memory and revise its own output. Moonshot describes one research report as passing through more than 120 rounds of recursive improvement.
Those features can look like learning, but they describe several different mechanisms.
A long context allows the model to reconsider earlier material during the same task.
External memory allows a product to store and retrieve information later. Moonshot’s API platform offers a memory tool for retaining conversation history and preferences.
Recursive improvement allows the model to critique and revise an output repeatedly.
Tool use allows the system to observe results and choose another action.
None of these necessarily changes the K3 model itself.
Continual learning normally requires a system to acquire knowledge or skills across a sequence of experiences while retaining useful previous capabilities. The component being changed might be the model weights, an external memory, a prompt, a skill library or the surrounding software. The mechanism must be identified before improved behaviour can be described as learning.
K3 is therefore better described as a powerful static foundation model capable of supporting stateful agents. A separate continual-learning system could be built around it, but Moonshot has not presented K3 itself as an online learner whose parameters improve automatically after deployment.
What limitations does Moonshot acknowledge?
Moonshot’s launch post identifies several weaknesses.
K3 was trained to preserve its previous reasoning history. If an agent framework does not pass that history back correctly, performance can become unstable. Switching from another model to K3 halfway through a session may also cause problems.
The company also says K3 can be excessively proactive. Because it is optimised for difficult, long-duration tasks, it may make unexpected decisions when instructions are ambiguous or a minor problem appears.
Developers using K3 in sensitive systems may therefore need to define tighter boundaries, permissions and approval stages.
Moonshot further acknowledges that K3’s overall user experience remains behind Claude Fable 5 and GPT-5.6 Sol, despite its competitive benchmark performance.
This is an important distinction. A model can score well while remaining less predictable, polished or convenient in ordinary use.
What is confirmed and what remains unclear?
| Confirmed | Unclear or unverified |
|---|---|
| K3 has 2.8 trillion total parameters. | The exact number active during ordinary inference. |
| It uses a sparse mixture-of-experts architecture. | How performance changes under different routing loads. |
| It supports a one-million-token context window. | Reliability across the entire context length. |
| It is available through Kimi products and the API. | When unrestricted new subscriptions will resume. |
| Independent evaluations place it near the frontier. | Its long-term position after broader testing. |
| Moonshot plans to release weights on 27 July. | The final licence and practical openness. |
| Its API costs $3 per million uncached input tokens and $15 per million output tokens. | Full task cost across different agent workflows. |
| Moonshot has disclosed several model limitations. | Complete safety, security and political-bias results. |
| The model can use tools and long-running agents. | Whether launch demonstrations can be independently reproduced. |
| External memory is available as a platform tool. | Any mechanism for continual updating of K3’s parameters. |
Who is affected?
Developers
K3 offers another credible option for coding agents, research systems and long-context applications. Developers must weigh performance against cost, infrastructure, licence terms and the risks created by excessive agent autonomy.
Open-model researchers
The scheduled weight release could provide an unusually large system for studying sparse routing, long-context behaviour, quantisation and agentic reasoning. Access may remain limited by the enormous hardware requirement.
American AI companies
K3 increases pricing and performance pressure on closed-model providers. Its strongest challenge may not be absolute benchmark leadership, but the possibility that developers can obtain comparable results from a model whose weights are eventually downloadable.
Chinese AI companies
Moonshot’s launch places pressure on domestic competitors as they race to release larger models, secure computing capacity and attract capital. Reuters reported sharp declines in the shares of some listed Chinese AI companies following K3’s announcement, although short-term market reactions should not be treated as a measure of lasting competitive advantage.
Businesses
Companies gain another model to test, but K3’s Chinese origin may introduce procurement, data-location, compliance and geopolitical questions in some jurisdictions. These should be assessed using the actual service terms and deployment arrangement rather than assumptions based solely on nationality.
Continual-learning developers
K3 could serve as a capable foundation model inside an adaptive system. Its release does not resolve the harder problem of how an AI system should learn repeatedly from production experience without forgetting, drifting or becoming less safe.
What happens next?
The most important immediate date is 27 July 2026, when Moonshot says it will release K3’s complete model weights and technical report.
That release should answer several questions:
What licence governs use and redistribution?
Will commercial use be permitted?
How much hardware is required for different quantisations?
Will the training and inference code be included?
How much of the benchmark suite can be reproduced?
Will Moonshot publish safety and security evaluations?
Can independent researchers confirm the architectural claims?
The subscription shortage is the second issue to watch. Moonshot must increase capacity without allowing latency, reliability or cost to undermine adoption.
The longer-term test will be sustained deployment.
Launch-week benchmarks can establish that a model is promising. They cannot show how reliably it performs across months of coding, enterprise workflows, different languages, adversarial prompts and changing real-world conditions.
Kimi K3 is not significant simply because 2.8 trillion is a large number.
Its importance comes from the combination of scale, sparse architecture, long context, multimodal input and credible performance on complex agent tasks. Independent evaluations suggest Moonshot has produced a system close to the strongest models available from much larger American competitors.
The qualifications are equally important.
K3’s weights are not yet public. Its final licence is unknown. Its full technical and safety documentation remains unpublished. Several launch demonstrations have not been independently reproduced, and the company has already encountered the operational difficulty of serving such a demanding model at scale.
K3 therefore represents both sides of the current AI race.
Model capability is spreading beyond a handful of American laboratories. The infrastructure required to deploy that capability remains highly concentrated, expensive and difficult to scale.
The release scheduled for 27 July will determine whether Kimi K3 becomes a genuinely accessible open-weight frontier model or a technically open system that only the largest organisations can afford to run.
Frequently asked questions
What is Kimi K3?
Kimi K3 is a large multimodal artificial intelligence model created by Beijing-based Moonshot AI. It contains 2.8 trillion total parameters, supports a one-million-token context window and is designed for coding, research, visual reasoning and long-running agent tasks.
When was Kimi K3 released?
Moonshot published its Kimi K3 announcement on 16 July 2026. The model became available through Kimi’s hosted products and API at launch. Moonshot says the downloadable weights and complete technical report will be released by 27 July 2026.
Is Kimi K3 open source?
The hosted version is not currently open source. Moonshot has promised to release the model’s full weights on 27 July, but the final licence and supporting materials have not yet been published. Whether K3 qualifies as broadly open source will depend on what is released and which uses the licence permits.
Is Kimi K3 better than ChatGPT or Claude?
No single benchmark can answer that universally. Independent evaluations place K3 near leading OpenAI and Anthropic models, with particular strength in coding and web-development tasks. It remains behind Claude Fable 5 and GPT-5.6 Sol on some overall evaluations, while leading selected tests.
Why does Kimi K3 have 2.8 trillion parameters?
K3 uses a mixture-of-experts architecture. The system contains a large pool of specialised expert components but selects only a small number for each token. This provides more total capacity without activating all 2.8 trillion parameters during every inference step.
Can Kimi K3 run on a laptop?
No practical consumer deployment has been demonstrated. Moonshot recommends infrastructure with at least 64 accelerators, placing self-hosting beyond ordinary laptops, workstations and most small companies. Smaller quantised versions may eventually reduce the requirement, but none had been released at publication.
How much does the Kimi K3 API cost?
Moonshot charges $3 per million uncached input tokens and $15 per million output tokens. Cached inputs cost $0.30 per million tokens. The cost of completing a task also depends on the amount of reasoning, context and output the model uses.
Why did Moonshot stop accepting new subscriptions?
Moonshot said demand during the first 48 hours exceeded forecasts and pushed its computing clusters close to their capacity limits. It temporarily paused new consumer subscriptions while prioritising existing paid users and adding infrastructure.
Does Kimi K3 learn continuously?
There is no public evidence that K3 automatically updates its model weights from each interaction. It can use long context, external memory, tools and repeated self-correction, but these mechanisms are not equivalent to a foundation model learning continuously after deployment.
Sources
https://www.kimi.com/blog/kimi-k3https://www.moonshot.cn/en/https://platform.kimi.ai/docs/pricing/chat-k3https://artificialanalysis.ai/models/kimi-k3https://artificialanalysis.ai/articles/kimi-k3-achieves-3-in-the-artificial-analysis-intelligence-index-comparable-to-opus-4-8-and-gpt-5-5/https://www.vals.ai/models/kimi_kimi-k3https://lmarena.ai/leaderboard/codehttps://www.reuters.com/world/china/chinas-moonshot-unveils-worlds-largest-open-ai-model-closing-us-rivals-2026-07-17/https://www.reuters.com/legal/transactional/chinas-moonshot-pauses-kimi-subscriptions-amid-hot-demand-ipo-push-2026-07-20/https://apnews.com/article/4c66a2e0f557ce79d3cc2d769c9a6226