Kimi K3, the first open AI model with 2.8 trillion parameters

What Kimi K3 is and why this launch matters

Moonshot AI introduced Kimi K3, its most capable model to date. It's a neural network with 2.8 trillion parameters, built on the company's own technologies called Kimi Delta Attention and Attention Residuals. The model works with images natively and holds up to one million tokens of text in memory at once, roughly the volume of several thick books.

The key word here is "open". Kimi K3 became the world's first open model at this scale, meaning its inner workings will be available to anyone rather than locked behind a company's closed access. The full model weights are set to be released by July 27, 2026.

The company openly admits that in overall performance, the model still trails the strongest closed solutions, Claude Fable 5 and GPT-5.6 Sol. Even so, Kimi K3 delivers frontier-level results across the full evaluation suite and consistently outperforms every other model it was tested against.

Why Kimi K3's size set a record

Kimi K3 became the first open model to reach 2.8 trillion parameters. It's the latest step in the company's sustained push at the scaling frontier, for nine of the past twelve months Kimi models have set the upper bound of open model sizes.

^{Open frontier model size over time. Total parameters of each company's flagship model, Jul 2025 - Jul 2026. Solid means released as of Jul 16, 2026, dotted means frontier held with no new release since, right label shows company and latest size. Only open models and flagship models are included.}

But size alone means little without solid architecture. Moonshot AI's engineers applied several technical solutions that let the model turn raw computing power into actual intelligence more effectively. Thanks to those changes, overall scaling efficiency improved roughly 2.5 times compared to the previous Kimi K2 model.

^{Kimi K3 architecture, the Stable LatentMoE and KDA modules on the left, the AttnRes operation at the top right, and the Block Attention Residuals backbone on the right}

How Kimi K3 handles coding

Kimi K3 is strong at long-running coding tasks. The model can operate with minimal human oversight, sustain long engineering sessions, navigate massive codebases, and orchestrate terminal tools.

Worth noting separately are tasks that mix coding with visual work. The model uses screenshots and visual material to improve game development, frontend interfaces, and 3D modeling.

Here's how Kimi K3 performs on the main coding benchmarks,

• On FrontierSWE, the model scores 81.2, behind only Claude Fable 5 at 86.6

• On ProgramBench, Kimi K3 takes first place with 88.8 points

• On SWE Marathon, the model also leads with a score of 42

• On the company's internal Kimi Code Bench 2.0, it places second at 72.9, behind Fable 5

So on certain coding tasks, an open model already beats the paid flagships, and on the rest it follows closely behind. For an open solution, that's a genuinely serious result.

Four examples of what Kimi K3 did on its own

The model's capabilities show best through the specific examples the company shared in its report.

The first involves optimizing code for graphics cards. The model was given up to 24 hours to figure out and speed up four different tasks on its own. Kimi K3 performed on par with Claude Fable 5 and clearly outperformed Opus 4.8, GPT-5.6 Sol, and GPT-5.5. On one task, after 15 hours of nonstop work, the model designed a novel two-phase algorithm and cut execution time from 283.6 to 114.4 milliseconds.

The second example is even more interesting. The model was asked to build an entire GPU programming system from scratch. Kimi K3 developed a compact compiler called MiniTriton that matches existing professional solutions in speed and even beats them on certain workloads.

The third example involves game development. Kimi K3 created a browser-based 3D martial arts role-playing game with melee combat, quests, inventory, dynamic weather, and interiors you can walk into. The model built the game environment itself, tuned object collisions, and designed the entire gameplay experience.

The fourth example is the most unexpected. As a proof of concept, Kimi K3 designed a computer chip to run a small model built on its own architecture. In a single autonomous 48-hour run, the model built, optimized, and verified the chip using open-source design tools. So an AI designed a chip for another AI, which says a lot about how long the model can stay on a single task.

How Kimi K3 helps with science and research

Kimi K3 bridges scientific literature and working code, autonomously implementing and validating complex computational research workflows.

In one example, the model completed in about two hours what would normally take an experienced researcher one to two weeks. To reproduce well-known relations from computational astrophysics, it reviewed and cross-validated more than 20 papers, implemented the full numerical pipeline, evaluated over 300 equations of state, spotted inconsistencies in published formulas, wrote more than 3,000 lines of Python code, and produced an interactive dashboard for exploring the results.

That kind of result shows exactly where the boundary of AI capability sits right now. The model doesn't replace the researcher, but it takes on a huge volume of routine yet technically demanding work.

What Kimi K3 does for knowledge work and research

Kimi K3 has made real progress on tasks that need to be carried through from start to finish. The company shared several examples of what the model produces in Kimi Work mode.

In the first case, the model built an interactive research website covering 42 years of the specialized chip industry. To prepare it, the model ran over 2,800 web searches, more than 1,100 terminal data pulls, and processed over 11,000 pages including 87 quarterly reports and 99 original documents. All of that went through more than 120 rounds of recursive self-improvement.

In the second example, the model produced a consulting-style industry report on controlled nuclear fusion with interactive visualizations, timelines, and slides. In the third, it analyzed 391 gravitational wave events using more than 20 concurrent subagents, producing 7 scientific visualizations, 2 tables, and a literature synthesis from over ten papers.

Two new features also appeared in Kimi Work. Widgets let you generate interactive components directly inside a chat and connect them to your own data or external plugins for continuous updates. The Dashboard gathers the widgets you care about most into one persistent view organized around a specific topic or project.

How Kimi K3 handles video and animation

Kimi K3 does well with animation and video editing because its architecture natively understands text, images, and video inside a single model rather than through separate add-ons.

In one example, the model created an animated explainer in a recognizable educational video style, walking through its own architecture. It translated technical ideas into animated diagrams and transitions on its own.

In another, Kimi K3 edited its own teaser video from 56 source clips. The model handled clip selection, motion-matched cuts, frame-accurate beat synchronization, audio processing, and several rounds of revisions. A dense short video like that typically takes an experienced editor one to two working days, or a beginner three to five.

How much Kimi K3 costs and where to try it

Kimi K3 is available today across several platforms. Here's exactly where,

• In the Kimi mobile app for iOS, Android, and HarmonyOS, and on the website

• In the Kimi Work desktop app, version 3.1.0 or later, for Windows and Apple silicon Macs

• In the terminal through Kimi Code, selecting the model with the /model command

• Through the API on the Kimi platform, where the model is listed as kimi-k3

API pricing looks like this, 0.30 dollars per million tokens on a cache hit, 3 dollars per million on a cache miss, and 15 dollars per million output tokens. Thanks to a specialized request-handling architecture, the company's official API achieves a cache hit rate above 90 percent on coding workloads, which means in practice most requests land at the lowest price tier.

At launch, Kimi K3 runs at maximum thinking effort by default, with lower-effort modes planned for future updates.

What limits Kimi K3 has

The company openly listed the model's weak spots, and they're worth knowing upfront.

The first limit involves thinking history. The model was trained in a mode where its full reasoning history is preserved. If the tool you're working through doesn't pass that history back in full, or if you switch to Kimi K3 midway through a session started with another model, output quality can drop sharply. The company recommends using verified tools like Kimi Code and avoiding mid-session model switches.

The second limit is excessive proactiveness. The model was trained on long, difficult tasks, so when it hits small ambiguities or vague instructions, it may make decisions on your behalf without asking. If you need the model to stay strictly within defined boundaries, spell out explicit constraints in your system prompt.

And third, the company itself admits that despite strong benchmark results, Kimi K3 still lags noticeably behind Claude Fable 5 and GPT-5.6 Sol in overall user experience. That's an honest note worth factoring in when picking a model for daily work.

An Open Model Holding Its Own

Kimi K3 highlights the biggest shift in the AI market, the gap between open and closed models has narrowed so much that calling it a meaningful lag no longer fits. A model with 2.8 trillion parameters beats paid solutions on certain coding tasks, designs chips autonomously, and assembles research reports that used to take a person weeks.

That said, a perfect model still doesn't exist. In overall usability, Kimi K3 trails the flagships from Anthropic and OpenAI, it has quirks you need to know about in advance, and for some tasks another model will be more accurate or cheaper. The only way to find out is on your own real task, not from someone else's scoreboard.

The problem is, registering separately on a Chinese platform, sorting out payment, and keeping other subscriptions running alongside just to compare takes too much time and money. It's far more practical when every current AI model sits in one place.

That's exactly what unitool.ai is for. Get one subscription and unlock access to the newest models at once, including fresh releases like Kimi K3, no foreign cards and no separate sign-up on each platform. See for yourself which model turns out stronger and more cost-effective for your specific tasks.