Open source AI is no longer sitting months behind the closed frontier models. That is the big story behind Moonshot AI’s Kimi K3, a massive model that is delivering shockingly capable results on long-horizon coding, research, 3D design, automation, and creative production tasks.
This is not another model that can merely summarize an email or generate a few lines of boilerplate. Kimi K3 was put through genuinely difficult, multi-step workflows: coding a liquid physics simulator from scratch, building an animated V8 engine in Blender, researching NVIDIA’s latest financial results and turning them into a narrated video, creating a working 3D piano, producing a playable shooter game, and conducting deep biomedical research.
The results are not flawless. Some of the tests revealed major limitations, especially in medical image recognition and visual search. But the broader takeaway is still enormous: a model expected to release its weights openly is now operating in the same conversation as leading proprietary models such as GPT 5.6 and Claude Fable.
For Canadian businesses, startup founders, IT leaders, developers, and creative teams, this matters right now. The ability to deploy, customize, and potentially run high-performance AI systems outside a closed vendor ecosystem could reshape how organizations across the GTA and beyond approach software development, knowledge work, research, and content production.
Table of Contents
- Kimi K3 intro
- Kimi code
- Liquid physics test
- Blender v8 engine
- Financial explainer video
- 3D model test
- Luma Agents
- Mecha shooter
- Music composition
- Cancer detection
- Find the frog
- Deep research
- Specs and benchmarks
- Open source
- FAQ
Kimi K3 intro
Kimi K3 is Moonshot AI’s latest large language model, built for more than ordinary chat. Its focus is long-horizon agentic work: tasks where the system needs to plan, use tools, inspect results, troubleshoot failures, and keep going until it reaches a usable outcome.
That distinction is crucial. Most leading AI models can already handle basic business tasks, including drafting emails, summarizing meetings, extracting notes, and writing short documents. The real frontier is whether an AI can take a complex objective and independently carry out dozens or hundreds of connected steps.
Kimi K3 appears designed for exactly that category of work. It can work through coding environments, browsers, local projects, files, APIs, research workflows, and creative applications. In practical terms, it acts less like a single-response chatbot and more like a persistent technical operator that can execute a plan.
There is a tradeoff: this kind of careful, iterative work can be slow. Several of the most demanding tasks took 30 to 40 minutes to complete and consumed millions of tokens. Still, the output quality on difficult assignments is what makes Kimi K3 so interesting.
Kimi code
Kimi K3 can be used through its online chat interface at Kimi.com, but that does not fully reveal what the model can do. The more powerful route is Kimi Code, Moonshot AI’s native agentic harness for working across projects, folders, and local files.
Kimi Code can operate on multiple persistent projects at once. A project may include many files and folders, allowing the model to work through a proper development environment rather than generate isolated snippets of code. This is where the concept of an “army of agents” starts to become real: different tasks can be organized into separate projects while the AI handles planning, creation, testing, and revisions.
Although it can be used in a terminal, the more accessible workflow is through Visual Studio Code. The installation process is straightforward:
- Open Visual Studio Code and go to Extensions.
- Search for Kimi Code.
- Install the extension and open the desired project folder.
- Select the latest K3 model in the extension panel.
- Choose a thinking mode, or use plan mode when you want strategy before execution.
Plan mode is especially useful for business technology teams. It allows an organization to inspect the model’s proposed approach before it touches source files, tools, or external services. That can help teams establish better oversight for AI-assisted development work.
Liquid physics test
The first major test was brutal by design: build a standalone HTML liquid splash simulator with adjustable gravity, viscosity, cohesion, lighting, glow, colour, rain effects, and webcam hand tracking. There was one major restriction: no Three.js and no external libraries.
That meant Kimi K3 had to create the physics and interaction logic itself rather than lean on a ready-made graphics framework. The system spent roughly 30 minutes planning, building, opening the result in a browser, taking screenshots, identifying defects, and fixing them.
The final application included an adjustable liquid simulation, controls for gravity and material-like behaviour, lighting tools, colour controls, rain generation, and hand-based interactions through a webcam. Lowering gravity caused the simulated liquid to float upward. Raising viscosity altered its behaviour toward thicker, more goo-like movement. Adjusting cohesion and surface detail changed the visual structure of the fluid.
This is a wildly impressive result for an open source model because the task combines front-end development, physics logic, browser interaction, computer vision input, user interface design, and self-testing. Kimi K3 used more than 2 million input tokens and 46,000 output tokens on this assignment, which illustrates both its persistence and its computational appetite.
Blender v8 engine
The Blender test pushed Kimi K3 into 3D automation. With Blender MCP enabled locally, the model was instructed to connect to the server and build a realistic animated V8 engine.
It took around 40 minutes. During that time, Kimi K3 created the engine components, built a cutaway setup, considered camera and lighting, and developed the animation. The final result included visibly separate mechanical elements and moving pistons in a fully animated 3D model.
The level of task coordination is the headline. Building a V8 engine is not simply a matter of generating shapes. It requires the model to organize a complex scene graph, create components, position them logically, and animate them in a believable way.
The cost in processing was substantial: approximately 17 million input tokens and 152,000 output tokens. That is clearly not a quick workflow for a casual experiment. But for product teams, industrial designers, engineering firms, and visualization specialists, this demonstrates a future where AI can accelerate early prototyping and technical visualization.
Canadian manufacturers and engineering companies should pay attention. The immediate opportunity is not to replace specialized designers. It is to compress the time needed to turn an idea, workflow, or early design brief into something visual, inspectable, and ready for expert review.
Financial explainer video
Kimi K3’s multi-tool capabilities were tested with a business-focused prompt: locate NVIDIA’s most recent quarterly financial report, analyze it, and create a roughly one-minute 16:9 explainer video. The output needed charts, graphs, visual elements, background music, animated sequences built with Hyperframes, and a Gemini TTS voiceover.
No direct financial-report link was supplied. Kimi K3 needed to find the source material itself, process the information, plan the script, generate the narration, build visuals, manage audio levels, and render a usable video.
The resulting explainer highlighted NVIDIA’s reported $81.6 billion in quarterly revenue, year-over-year growth of 85%, $75.2 billion in data-centre revenue, and $91 billion in next-quarter guidance. It also included references to free cash flow, profitability, networking growth, shareholder returns, and the company’s stated China assumption in its data-centre guidance.
This is where Kimi K3 becomes especially relevant for finance, investor relations, consulting, and corporate communications teams. An organization could potentially move from a source document to an internal presentation asset, narrated briefing, or executive summary far faster than with a conventional manual workflow.
That does not eliminate the need for financial review. It makes review more important. AI can accelerate the assembly of a polished narrative, but executives must validate data, wording, assumptions, and compliance requirements before publication or investor use.
3D model test
The 3D model test asked Kimi K3 to build a beautiful interactive grand piano in a single HTML file. The piano needed a full 88-key layout, playable notes, animated internal mechanics, a slider for a smooth exploded view, and a demonstration performance of Beethoven’s Für Elise.
Kimi K3 created a functioning piano with keys that could be clicked or played through glissando-style dragging. The piano could move between a fully assembled presentation and an exploded configuration that revealed strings, hammers, dampers, and other internals.
There was one visible flaw in the first result: the piano lid was not positioned properly and fell to the ground rather than resting on its support. A single follow-up instruction corrected it. That is an important real-world point. Even when a model produces an extraordinary first pass, iterative feedback remains part of the workflow.
The fact that this project used only about 400 input tokens and 5,000 output tokens makes it even more striking. For interactive product demonstrations, educational tools, digital showrooms, and technical explainers, this kind of browser-based 3D work could become highly valuable.
Luma Agents
Creative teams that repeat the same workflow again and again face a different problem: not how to generate one asset, but how to standardize excellent results across a team. Luma Agents, from Luma AI, is positioned as a system for turning creative processes into reusable AI workflows.
The core idea is simple. Upload references, describe a desired output in natural language, and let the agent create the workflow. Once it works, save it as a reusable Luma Skill.
Examples include:
- Turning a reference image into a complete brand kit and colour palette.
- Generating a fashion-show video from a model image and clothing reference.
- Saving a successful workflow so a team can apply it to new assets later.
- Bundling several skills into a shareable creative toolkit.
The business implication is huge. Marketing teams can reduce creative inconsistency. Agencies can package repeatable production processes. Internal design teams can share systems rather than relying on undocumented individual expertise.
For Canadian companies scaling content operations, reusable AI workflows may become as important as reusable code. The competitive edge is not just generating images or videos. It is operationalizing the process so the same quality can be delivered repeatedly.
Mecha shooter
Game development is one of the areas where Kimi K3 looks particularly strong. The task was to create a third-person 3D shooter using Three.js, with a futuristic mecha warrior fighting waves of alien creatures from the ground and sky. It also needed publicly available 3D assets and a polished AAA-style look.
This is exactly the sort of assignment that often breaks AI coding systems. It requires asset discovery, importing, game-state management, camera controls, player movement, enemy behaviour, combat effects, user interface instructions, and browser verification.
Kimi K3 entered plan mode, outlined its approach, found assets, built the game, and then opened the browser to inspect the final experience. It verified elements such as the mecha’s movement, jet glow, muzzle flashes, and combat presentation.
The completed game featured keyboard movement, mouse aiming, a plasma cannon, jump jets, and a dash mechanic. It reportedly looked substantially stronger than comparable results produced by Claude Fable and GLM 5.2 in this type of test.
This took roughly 12 million input tokens and 135,000 output tokens. It is another reminder that Kimi K3 is not necessarily the fastest option. But for teams exploring interactive prototypes, simulations, training experiences, or game concepts, the quality ceiling is becoming very serious.
Music composition
Kimi K3 was also asked to create a digital audio workstation-style interface featuring piano, synth pluck, strings, drums, and bass. Each instrument needed a piano roll for drawing notes. The goal was a 32-bar song with effects, panning, automation, and mastering controls.
The first attempt created a reasonable musical foundation but lacked enough complexity. A follow-up prompt requested greater variation, more instruments, risers, drops, wider sound, and volume and effects automation.
The refined result included reverb, delay, track automation, risers, crashes, and a more layered composition. It was not a professional, Grammy-level production, but it was a surprisingly good piece of music generated from scratch without conventional virtual instruments.
This test required about 1.5 million input tokens and 35,000 output tokens. The broader lesson is that AI-generated creative interfaces can now combine software design with content creation. That opens possibilities for music education, rapid sound design, interactive media, and prototype creation, even if human artists remain essential for final professional work.
Cancer detection
Here is where the excitement needs a hard reality check. Kimi K3 was given six medical scans, each containing a type of tumour, and asked to identify the tumours.
The model used Python tools to zoom into individual image panels and investigate suspicious areas. It correctly identified one image as a meningioma. However, it got the other five cases wrong.
That one-out-of-six result was reportedly better than GPT 5.6, which identified none of the tumours, while Claude Fable declined to answer biomedical questions. But “better than competing models” is not remotely the same as clinically reliable.
Kimi K3 should not be used to diagnose cancer from medical scans. The test demonstrates that the model can inspect images, use tools, and attempt visual reasoning. It also demonstrates why healthcare AI requires rigorous validation, regulatory oversight, clinical expertise, and extremely cautious deployment.
For Canadian health organizations, this is a reminder that advanced AI can support research and administrative analysis, but patient-facing diagnostic claims demand a dramatically higher standard of evidence.
Find the frog
The “find the frog” challenge is a deceptively simple visual reasoning test. An image containing a hidden frog among leaves was uploaded with a general question asking whether any animal could be found and circled.
Kimi K3 went to work. It divided the image into sections, zoomed in repeatedly, created snapshots, adjusted sharpness and contrast, and even applied a spectral residual saliency technique to identify potentially important regions.
After all that work, it concluded that the image contained only dead leaves and no animal. The frog was missed.
This is a fantastic example of why organizations should not confuse elaborate reasoning traces with correctness. Kimi K3 demonstrated persistence and sophisticated tool use, but it still failed the actual objective. Strong AI systems can make remarkably capable mistakes, sometimes after a great deal of analysis.
Deep research
Kimi K3 Max includes a deep research option through the online interface. To test it, the model was asked to examine propagation mechanisms in Alzheimer’s disease, compare monoclonal antibody therapies targeting different proteins, and provide tables, visualizations, and scientific references.
The task required a deep literature search, synthesis of complex biomedical evidence, creation of figures, and a structured report. It took approximately 20 minutes.
The final research output included:
- An executive summary.
- Detailed sections on amyloid beta propagation.
- Detailed sections on tau propagation.
- Flow charts and explanatory figures.
- Tables comparing antibody therapies.
- Visual summaries of trial results.
- A critical appraisal of Phase 3 evidence.
- A final synthesis and outlook section.
The report was described as thorough, information-dense, and supported by citations. This is where Kimi K3’s writing style may appeal to research-intensive organizations. Compared with shorter, more concise outputs from some models, Kimi K3 appears willing to produce a substantial, structured report.
Still, deep research outputs must be independently checked. Citations, study interpretations, tables, and clinical conclusions all require expert review. AI can help research teams assemble a first-pass map of a complex topic, but it cannot replace the scrutiny of scientists, analysts, or domain experts.
Specs and benchmarks
Kimi K3 is reportedly a 2.8 trillion parameter model built on Moonshot AI’s Kimi Delta Attention and Attention Residuals architecture. It has native vision capabilities and a context window of up to 1 million tokens.
One million tokens is roughly equivalent to 700,000 words. In practical terms, that could cover a small to medium-sized codebase or a very large collection of business documents. For organizations handling long technical specifications, contracts, knowledge bases, financial records, or software repositories, large context windows are a major capability.
Benchmark results place Kimi K3 unusually close to leading proprietary models. Across software engineering and agentic task measures, it reportedly approaches Claude Fable 5 and GPT 5.6, while significantly outperforming earlier open models such as GLM 4.2 and GLM 5.2.
Highlights cited for Kimi K3 include strong results across:
- DeepSWE software engineering evaluations.
- TerminalBench and Frontier SWE agentic coding tasks.
- ProgramBench and other development benchmarks.
- General agentic work, including AutomationBench and SpreadsheetBench.
- BrowseComp and visual-agent tasks.
- Chart analysis and scientific figure interpretation.
- Code Arena front-end web development tests.
Cost efficiency is another major differentiator. On cost-versus-performance charts, Kimi K3 lands toward the desirable upper-left area: high task performance at a lower cost per task than major proprietary competitors. For enterprise AI leaders, that matters as much as raw intelligence. A system that is excellent but impossible to scale economically is not a complete business solution.
Kimi K3 also recorded a 51% hallucination rate on one cited leaderboard, lower than Claude Fable 5 at 55% and GPT 5.6 at 89% in that test. That does not mean hallucination is solved. A 51% figure is still a giant warning label. It simply suggests that Kimi K3 may be less prone to this issue than certain competitors under that particular evaluation.
Open source
The most disruptive part of the Kimi K3 story is not only its performance. It is the plan to release the model weights openly, with the full weights stated as due by July 27.
That could give developers and enterprises more control over how they use advanced AI. Instead of relying exclusively on a closed API provider, organizations may gain more flexibility around deployment, customization, integration, and experimentation.
Kimi K3 is available through several routes:
- Kimi.com for the native online chat interface.
- Kimi Work, a desktop app designed to work with local files.
- Kimi Code through Visual Studio Code for agentic software workflows.
- API access for developers building their own products and systems.
For Canadian technology leaders, the strategic question is no longer whether open source AI can be useful. It is whether their organization has a plan to evaluate it safely. The tools are getting more powerful, more affordable, and more capable of acting across real workflows.
Kimi K3 is not perfect. It can be slow. It can miss obvious details. It can be wrong on high-stakes questions. But its ability to plan, create, test, iterate, and deliver surprisingly sophisticated outputs marks a serious shift in the AI market.
The open source AI frontier is moving at an insane pace. Companies that wait for complete certainty may discover that competitors have already built the internal expertise, agentic workflows, and AI operating models needed to move faster.
FAQ
What is Kimi K3?
Kimi K3 is a large AI model from Moonshot AI designed for long-horizon agentic tasks, including coding, research, tool use, document processing, creative work, and software development.
Is Kimi K3 open source?
Moonshot AI has stated that Kimi K3 will be open sourced, with full model weights planned for release by July 27.
How can developers use Kimi K3?
Developers can access Kimi K3 through Kimi.com, the Kimi Work desktop application, Kimi Code in Visual Studio Code, and the model’s API.
Can Kimi K3 diagnose cancer from scans?
No. In a six-image tumour identification test, Kimi K3 correctly identified only one case. It should not be used for medical diagnosis or clinical decision-making.
Why should Canadian businesses care about Kimi K3?
Kimi K3 suggests that powerful open source AI may become a viable option for Canadian organizations seeking lower-cost, customizable systems for software development, research, content creation, technical visualization, and agentic business workflows.
Is your organization prepared to test open source AI models that can plan, code, research, and create across entire workflows? The AI advantage is rapidly becoming an execution advantage.
