Moonshot AI has introduced one of the largest openly announced artificial intelligence models yet—but the most important part of the release is still several days away.
Chinese artificial intelligence company Moonshot AI has unveiled Kimi K3, a new frontier model built around an enormous 2.8 trillion parameters, native multimodal capabilities, and a 1 million-token context window.
The company is positioning Kimi K3 as an advanced system for long-horizon programming, research, document-heavy knowledge work, visual reasoning, and autonomous tool use. It is available now through Kimi’s consumer products, desktop workspace, coding agent, and developer API.
However, Kimi K3 is not yet a downloadable open-weight model in the practical sense. Moonshot says the full model weights and additional technical details will be released by July 27, 2026. Until then, developers can test the hosted version, but they cannot independently inspect or deploy the complete model.
Kimi K3 Reaches 2.8 Trillion Parameters
The headline specification is difficult to ignore. At 2.8 trillion total parameters, Kimi K3 is considerably larger than Moonshot’s earlier Kimi K2 family, which used a 1-trillion-parameter Mixture-of-Experts architecture.
Parameter count alone does not determine intelligence, speed, or output quality. It does, however, provide a sense of the model’s scale and the engineering challenge involved in training and serving it.
Kimi K3 uses a highly sparse Mixture-of-Experts, or MoE, architecture. Instead of activating the entire model for every token it generates, the system routes each token through a smaller selection of specialized components.
Moonshot says Kimi K3 contains 896 experts and effectively activates 16 experts at a time. The company has not yet disclosed a simple active-parameter figure comparable to the 32 billion active parameters advertised for the original Kimi K2, so those older numbers should not be applied to K3.
The architecture also introduces what Moonshot calls Stable LatentMoE, which is intended to improve routing and training stability as the number of experts increases.
A 1 Million-Token Context Window
Kimi K3 supports a native context window of up to 1 million tokens, giving it the theoretical ability to process very large collections of source code, reports, documentation, research papers, images, and conversation history within a single working session.
That does not mean every million-token prompt will be fast, inexpensive, or equally reliable. Long-context models can still overlook details, place uneven attention on different sections, or become less consistent as a session expands.
Still, the expanded context window could be useful for tasks such as reviewing large software repositories, comparing extensive document collections, tracing dependencies across long projects, and operating agents that need to retain substantial working history.
Moonshot reports that Kimi K3 reached a score of 90.4 on BrowseComp when evaluated with the full 1-million-token window and no additional context-management system. That result comes from Moonshot’s own evaluation disclosure and should be treated as a company-reported benchmark until it is replicated more broadly.
New Attention and Routing Techniques
Kimi K3 is built around two architectural technologies Moonshot calls Kimi Delta Attention and Attention Residuals.
Kimi Delta Attention is designed to reduce some of the memory and computational pressure associated with conventional attention mechanisms, particularly when processing very long sequences. Attention Residuals changes how the model retrieves information from earlier layers, selectively drawing from previous representations instead of simply accumulating them in the traditional manner.
Moonshot says the combination of these techniques, its expanded MoE design, and updated training methods produced an approximately 2.5-times improvement in scaling efficiency compared with Kimi K2.
That is an internal measurement rather than an independently established industry figure. The forthcoming technical report should provide more information about how Moonshot defined and calculated the improvement.
Kimi K3 Is Built for Long-Horizon Agentic Work
Moonshot is presenting Kimi K3 less as a conventional chatbot and more as an agent capable of working through extended, multi-stage projects.
The company says the model can navigate large software repositories, use terminal tools, work with visual feedback, generate and test code, and continue engineering tasks with relatively little human supervision.
In one company demonstration, Kimi K3 reportedly developed a compact GPU compiler called MiniTriton, including an intermediate representation, optimization passes, and a PTX code-generation pipeline. Moonshot says the resulting system matched or exceeded Triton on some supported tests.
Another demonstration involved an autonomous 48-hour run in which the model used open-source electronic-design-automation tools to produce and verify a small chip design intended to serve a compact neural model.
These demonstrations are technically interesting, but they remain curated examples published by the model developer. They do not prove that Kimi K3 can reliably complete every similarly complex project without supervision.
How Its Benchmark Position Should Be Interpreted
Moonshot says Kimi K3 performs competitively with leading proprietary systems across several coding, research, productivity, and multimodal evaluations. Independent trackers cited by Reuters also place it near the front of the current model market in selected tests.
Reuters reported that Arena.ai ranked Kimi K3 first in an evaluation focused on building web interfaces, while Vals AI placed it second overall behind Anthropic’s Claude Fable 5 in its comparison. Artificial Analysis reportedly found performance broadly comparable with several leading systems on complex multi-step work.
Those results are encouraging, but they do not establish a universal ranking. Different models may be evaluated through different agent frameworks, reasoning settings, tools, context-management systems, and execution environments.
Moonshot’s own benchmark notes acknowledge these differences. Some models were tested through Kimi Code, some through Claude Code, and others through OpenAI’s Codex harness. Several results also came from separate organizations rather than one standardized evaluation run.
The more defensible conclusion is that Kimi K3 appears to be a serious frontier competitor, especially for coding and long-horizon agentic work—not that it has conclusively defeated every closed model.
Open Model, but the Weights Are Not Available Yet
Moonshot describes Kimi K3 as an open model and says the full weights will be released by July 27. As of the initial announcement, however, the K3 weights were not present in Moonshot’s public Hugging Face catalog.
The company also had not published the complete technical report or clearly identified the final weight license in its launch article.
That distinction matters. A model can be announced as open before developers know exactly what they will be permitted to do with it. Commercial usage rights, redistribution terms, required attribution, geographic restrictions, and acceptable-use conditions can all depend on the final license.
Businesses should therefore avoid assuming that Kimi K3 will use Apache 2.0, MIT, or the same licensing structure as an earlier Moonshot release. The actual repository and license text should be reviewed once the weights are published.
Running Kimi K3 Locally Will Not Be a Consumer Project
The eventual release of the weights should not be confused with easy local deployment.
Moonshot says Kimi K3 uses quantization-aware training with MXFP4 weights and MXFP8 activations. Even at roughly four bits per weight, storing 2.8 trillion parameters would require approximately 1.4 terabytes before accounting for additional runtime memory, model metadata, caches, routing overhead, activations, and system reserves.
The company recommends deploying Kimi K3 on supernode configurations containing 64 or more accelerators. That places the full model firmly in the data-center category rather than the desktop, workstation, or conventional homelab market.
A single RTX 4090, RTX 5090, Intel Arc Pro B60, or consumer AI PC will not run the complete Kimi K3 model. Even aggressively compressed community versions would require substantial compromises and unusually large amounts of system memory.
For most developers and smaller businesses, realistic access will come through the official API or third-party inference providers rather than local hardware.
Kimi K3 API Pricing
Moonshot lists the initial Kimi K3 API rates as:
- $0.30 per million tokens for cached input
- $3 per million tokens for uncached input
- $15 per million tokens for output
The large difference between cached and uncached input reflects the importance of reusing prompt context. Coding agents and long-running work sessions often send substantial portions of the same repository, documentation, or conversation history with each request.
Moonshot says its Mooncake inference architecture produces a cache-hit rate above 90% in coding workloads on its own platform. Actual costs will depend heavily on whether a particular application can achieve similar context reuse.
Important Limitations Disclosed by Moonshot
Moonshot’s launch material includes several unusually direct warnings about Kimi K3’s behavior.
The company says K3 can become unstable when an agent framework fails to preserve and return its full reasoning history. Switching to K3 midway through a session started with another model may also reduce output quality.
Moonshot additionally warns that the model can be excessively proactive. Because it is trained for difficult, long-running assignments, it may make unexpected decisions when instructions are ambiguous or when it encounters relatively minor obstacles.
Applications that require strict boundaries should therefore use explicit system instructions, permission controls, sandboxing, spending limits, and human approval gates.
Moonshot also concedes that Kimi K3 still trails the overall user experience of the strongest proprietary models it evaluated. That is an important qualification for a release that will inevitably attract exaggerated claims based on parameter count alone.
Kimi K3 matters because it extends the open-model race into a scale previously associated almost entirely with closed, hyperscaler-operated systems. But its importance should not be confused with accessibility. A 2.8-trillion-parameter model may be downloadable after July 27, yet operating the complete system will remain beyond the reach of nearly every individual developer and small business. Its near-term impact will come primarily from lower-cost hosted inference, research access, community inspection, and the pressure it places on proprietary model providers—not from people running it on gaming PCs.
What Kimi K3 Means for the AI Market
Kimi K3 demonstrates how rapidly Chinese AI companies are moving from competitive smaller models toward systems intended to operate at the same frontier as the largest Western platforms.
It also highlights a growing split in the meaning of “open AI.” The weights may be publicly downloadable, but the resources required to operate them can still make access dependent on cloud platforms, specialist inference providers, and organizations with large compute clusters.
Even so, releasing the weights would allow researchers to examine the architecture, develop new inference methods, create specialized fine-tunes, test safety behavior, and potentially produce smaller distilled models derived from K3.
That is where the broader effect may emerge. The full 2.8-trillion-parameter system is unlikely to appear in a typical local AI setup, but the techniques and smaller derivatives it produces could eventually influence models that do.
Bottom Line
Moonshot AI’s Kimi K3 is an important release built around extraordinary scale: 2.8 trillion parameters, 896 experts, native multimodal processing, and a 1-million-token context window.
Its early coding and agentic results suggest that it belongs in the frontier-model conversation, although many of the most impressive figures still come from Moonshot’s own evaluations and demonstrations.
The most important test begins on July 27, when Moonshot says it will publish the weights and additional technical information. That release should reveal the model’s actual license, practical deployment requirements, framework support, and how quickly independent researchers can reproduce the company’s claims.
Until then, Kimi K3 should be viewed as a promising and potentially significant open-model announcement—not yet as a downloadable model that has already rewritten the economics of artificial intelligence.
 |
Local AI Hardware Demands Surge as Chinese MoE Models Challenge U.S. Open-Weight Standards
|
 |
Apple’s M7 Ultra Could Bring 1.5TB of Unified Memory to Local AI
|