Chinese AI company Moonshot AI has launched Kimi K3, its most capable model so far and the world’s first open 3T-class AI model.
Kimi K3 has 2.8 trillion parameters, native vision capabilities, and a 1 million-token context window. Moonshot says the model is designed for long-horizon coding, knowledge work, reasoning, and multimodal tasks.
The company says Kimi K3 still trails the strongest proprietary models, including Claude Fable 5 and GPT-5.6 Sol, in overall performance and user experience. However, it claims Kimi K3 delivers frontier-level results across its evaluation suite and consistently outperforms other tested open models.
| Category | Details |
|---|---|
| Model Name | Kimi K3 |
| Developer | Moonshot AI |
| Model Type | Open 3T-class AI model |
| Total Parameters | 2.8 trillion |
| Context Window | 1 million tokens |
| Native Vision | Yes |
| Architecture | Kimi Delta Attention, Attention Residuals, Stable LatentMoE |
| MoE Setup | 896 experts, 16 activated during use |
| Main Focus Areas | Long-horizon coding, reasoning, knowledge work, research, vision, video editing |
| Thinking Mode at Launch | Max thinking effort by default |
| Future Thinking Modes | Low and high effort modes planned later |
| Weights Release Date | Full model weights expected by July 27, 2026 |
| Technical Report | Coming later with architecture, training, and evaluation details |
Kimi K3 is built on Kimi Delta Attention and Attention Residuals.
Moonshot says these two architecture changes improve how information flows across long sequences and deep model layers. The model also uses Stable LatentMoE, which activates 16 out of 896 experts during inference.
According to Moonshot, these changes helped deliver about 2.5 times better scaling efficiency compared with Kimi K2. The company says the model can convert compute into intelligence more effectively because of these architecture and training improvements.
Kimi K3 is designed for long coding tasks with limited human oversight.
Moonshot says the model can work through large code repositories, use terminal tools, run long engineering sessions, and combine coding with visual reasoning. The company says this makes it useful for software engineering, frontend development, game development, CAD, and scientific programming.
In one example, Moonshot said Kimi K3 built a browser-based 3D open-world game using Three.js, WebGPU, and GPU compute. The project included a procedural environment with forests, a village, snowy mountains, dynamic weather, and externally generated character assets.
The company also said Kimi K3 created MiniTriton, a compact Triton-like compiler with its own tile-level IR layer, MLIR-based optimization passes, and a PTX code-generation pipeline.
| Benchmark | Kimi K3 | Claude Fable 5 | GPT-5.6 Sol | Claude Opus 4.8 | GPT-5.5 | GLM-5.2 |
|---|---|---|---|---|---|---|
| DeepSWE | 67.5 | 70.0 | 73.0 | 59.0 | 67.0 | 46.2 |
| Program Bench | 77.8 | 76.8 | 77.6 | 71.9 | 70.8 | 63.7 |
| Terminal Bench 2.1 | 88.3 | 84.6 | 88.8 | 84.6 | 83.4 | 82.7 |
| FrontierSWE | 81.2 | 86.6 | 71.3 | 66.7 | 64.9 | 67.3 |
| SWE Marathon | 42.0 | 35.0 | 39.0 | 40.0 | 14.0 | 13.0 |
| PostTrain Bench | 36.6 | 41.4 | 34.6 | 34.1 | 28.4 | 34.3 |
| MLS Bench | 48.3 | 49.9 | 46.2 | 42.8 | 35.5 | 40.4 |
| Kimi Code Bench 2.0 | 72.9 | 76.9 | 64.8 | 71.7 | 69.0 | 64.2 |
Moonshot says Kimi K3 performed strongly in GPU kernel optimization tests.
The company tested models in identical sandboxes for up to 24 hours across four tasks involving AttnRes, KDA, MLA, and GPGPU workloads. Moonshot says Kimi K3 performed competitively with Claude Fable 5 and substantially outperformed Claude Opus 4.8, GPT-5.6 Sol, and GPT-5.5 in that specific test.
In one AttnRes task, Kimi K3 reportedly optimized a training-side operation and cut forward-plus-backward time from 283.6ms to 114.4ms across 15 hours of iterations.
Moonshot also highlighted research coding. In one case, Kimi K3 reproduced the I–Love–Q universal relations in computational astrophysics. The company says the model reviewed and cross-validated more than 20 papers, evaluated more than 300 equations of state, generated over 3,000 lines of Python code, and produced an interactive HTML dashboard.
Kimi K3 is also aimed at agentic knowledge work.
Moonshot says the model can produce research reports, interactive visualizations, consulting-style documents, financial analysis, scientific dashboards, and editable presentations. It also introduced two new Kimi Work features: Widgets and Dashboard.
Widgets allow users to generate interactive components inside a chat. Dashboard keeps selected widgets in a persistent project or topic view.
| Benchmark | Kimi K3 | Claude Fable 5 | GPT-5.6 Sol | Claude Opus 4.8 | GPT-5.5 | GLM-5.2 |
|---|---|---|---|---|---|---|
| GDPval-AA v2 | 1668.0 | 1760.0 | 1748.0 | 1600.0 | 1494.0 | 1514.0 |
| BrowseComp | 91.2 | 88.0 | 90.4 | 84.3 | 84.4 | — |
| DeepSearchQA | 95.0 | 94.2 | — | 93.1 | — | — |
| Toolathlon-Verified | 73.2 | 77.9 | 74.9 | 76.2 | 73.5 | 59.9 |
| MCP Atlas | 84.2 | 84.7 | 83.6 | 83.6 | 82.8 | 82.6 |
| Automation Bench | 30.8 | 29.1 | 29.7 | 27.2 | 22.7 | 12.9 |
| Job Bench | 52.9 | 57.4 | 46.5 | 48.4 | 38.3 | 43.4 |
| AA-Briefcase | 1548.0 | 1583.0 | 1495.0 | 1354.0 | 1158.0 | 1260.0 |
| APEX-Agents | 37.6 | 43.3 | 39.9 | 39.4 | 38.5 | 35.6 |
| Office QA Pro | 63.3 | 69.9* | 63.2* | 63.9* | 60.9* | 41.4 |
| SpreadsheetBench 2 | 34.8 | 34.7* | 32.4* | 31.6* | 29.1* | 28.1 |
| DECK-Bench | 73.5 | 73.0 | 74.7 | 66.9 | 68.2 | 68.6 |
Kimi K3 also posted strong results in reasoning and knowledge benchmarks, though Moonshot’s own table shows Claude Fable 5 and GPT-5.6 Sol ahead in some areas.
| Benchmark | Kimi K3 | Claude Fable 5 | GPT-5.6 Sol | Claude Opus 4.8 | GPT-5.5 | GLM-5.2 |
|---|---|---|---|---|---|---|
| GPQA-Diamond | 93.5 | 92.6 | 94.1 | 91.0 | 93.5 | 91.2 |
| HLE-Full | 43.5 | 53.3 | 44.5 | 49.8* | 41.4* | — |
| HLE-Full With Tools | 56.0 | 63.0 | 58.0 | 57.9* | 52.2* | — |
Kimi K3 includes native multimodal support, meaning it can work with text, images, and video inside the same model.
Moonshot says this helps the model handle motion design, animation, video editing, visual reasoning, document understanding, and scientific visualization. In one example, the company said Kimi K3 created a 3Blue1Brown-style motion-graphics explainer of its own architecture.
In another example, the model reportedly edited its own teaser video from 56 source clips. Moonshot says it handled clip selection, motion-matched cuts, beat synchronization, audio processing, and several rounds of revision.
| Benchmark | Kimi K3 | Claude Fable 5 | GPT-5.6 Sol | Claude Opus 4.8 | GPT-5.5 | GLM-5.2 |
|---|---|---|---|---|---|---|
| MMMU-Pro | 81.6 | 81.2 | 83.0 | 78.9 | 81.2 | — |
| MMMU-Pro With Python | 83.4 | 86.5 | 84.6 | 82.7 | 83.2 | — |
| CharXiv RQ | 84.8 | 88.9 | 84.6 | 80.5 | 84.1 | — |
| CharXiv RQ With Python | 91.3 | 93.5 | 89.1 | 89.9 | 89.0 | — |
| MathVision | 94.3 | 94.8 | 95.8 | 86.7 | 92.2 | — |
| MathVision With Python | 97.8 | 98.6 | 97.8 | 97.1 | 96.8 | — |
| BabyVision With Python | 85.7 | 90.5 | 88.9 | 81.2 | 83.6 | — |
| ZeroBench Main Pass@5 | 23.0 | 23.0 | 17.0 | 17.0 | 22.0 | — |
| ZeroBench Main With Python Pass@5 | 41.0 | 46.0 | 35.0 | 34.0 | 41.0 | — |
| WorldVQA ForceAnswer | 51.0 | 56.7 | 41.8 | 39.1 | 38.5 | — |
| OmniDocBench | 91.1 | 89.8 | 85.8 | 87.9 | 89.4 | — |
| PerceptionBench | 58.5 | 57.2 | 59.7 | 47.2 | 55.8 | — |
Kimi K3 is available today through Kimi.com, Kimi Work, Kimi Code, and the Kimi API.
Moonshot says users can access Kimi K3 through the latest Kimi mobile app on iOS, Android, and HarmonyOS. Kimi Work is available through desktop app version 3.1.0 or later for Windows and Apple silicon Macs. Developers can use Kimi Code in the terminal and select Kimi K3 through the /model command.
| Platform | Availability |
|---|---|
| Kimi.com | Available now |
| Kimi App | Available on iOS, Android, and HarmonyOS |
| Kimi Work | Available through desktop app version 3.1.0 or later |
| Kimi Code | Available in terminal through the /model command |
| Kimi API | Available through Kimi API Platform |
| Kimi Enterprise | Available for organizations with enterprise-grade privacy and member management |
| API Pricing | Cost |
|---|---|
| Cache-hit input | $0.30 per million tokens |
| Cache-miss input | $3.00 per million tokens |
| Output | $15.00 per million tokens |
Moonshot says the official Kimi API is powered by Mooncake’s disaggregated inference architecture and can achieve a cache hit rate above 90 percent in coding workloads.
The full model weights are not available yet.
Moonshot says Kimi K3 is available through its products and API now, while the full model weights will be released by July 27, 2026. The company also plans to publish more details about the model’s architecture, training process, and evaluations in a technical report.
At launch, Kimi K3 uses maximum thinking effort by default. Low-effort and high-effort modes will be added in later updates.
| Category | Details |
|---|---|
| Biggest Strength | Long-horizon coding, large-context reasoning, knowledge work, and native multimodal tasks |
| Coding Strength | Strong results in Program Bench, Terminal Bench 2.1, FrontierSWE, and SWE Marathon |
| Knowledge Work Strength | Can produce interactive reports, dashboards, visual research outputs, and editable presentations |
| Vision Strength | Strong performance across document, chart, math-vision, and multimodal benchmarks |
| Video Capability | Can handle clip selection, editing, motion design, beat synchronization, and repeated revisions |
| Main Limitation | Still trails Claude Fable 5 and GPT-5.6 Sol in overall user experience |
| Thinking History Issue | Performance may become unstable if the harness does not preserve full thinking history |
| Proactiveness Issue | May make unexpected decisions on vague tasks unless strict boundaries are provided |
| Deployment Challenge | Large 2.8T-parameter scale means local deployment will require expensive hardware |
| Best Use Case | Developers, researchers, enterprises, and teams needing long-context agentic workflows |
Moonshot is presenting Kimi K3 as a major open-model release, not as a model that beats every closed-source rival.
The company says the model remains behind Claude Fable 5 and GPT-5.6 Sol in overall user experience. It also warns that Kimi K3 may become unstable if an agent harness does not preserve full thinking history, or if a session switches to K3 midway from another model.
Moonshot also says Kimi K3 may behave too proactively in some tasks. Since the model is trained for long-horizon work, it may make unexpected decisions when the user’s intent is vague. The company recommends adding clear boundaries in system prompts or project instructions.
Even with those limits, Kimi K3 is one of the most ambitious open AI releases so far. Its 2.8 trillion-parameter scale, 1 million-token context window, native vision support, and strong coding benchmarks make it a major new entry in the global race between open and closed AI models.
Get the latest tech news, telecom insights, and product launches wherever you prefer.
Add ProPakistani to Preferred Sources and see more of our stories in Google Search and Top Stories.