Table of Contents
Kimi K3 has taken over AI conversations this week, and for good reason. Released by Chinese startup Moonshot AI on July 17, 2026, it is a 2.8 trillion parameter open weight model that jumped straight into the top tier of global AI rankings within hours of launch.
If it has shown up in your news feed, your LinkedIn timeline, or a coding forum recently, you are not imagining the buzz. This is both a news story and an explainer.
Beyond covering why Kimi K3 launched to so much attention, this article breaks down how it actually works under the hood, how it compares to Claude and GPT models, what it costs, and why its release is being called one of China’s biggest AI breakthroughs yet.
What Is Kimi K3?
Kimi K3 is a large language model built by Moonshot AI, a Beijing- based startup backed by Alibaba. It is the newest entry in the Kimi model family, following Kimi K2 and Kimi K2 Thinking, both of which had already earned a reputation for strong coding performance at a low price. It raises the ambition considerably. With 2.8 trillion parameters, Moonshot describes it as the world’s largest open-weight model released to date, and the company positions the model as a genuine rival to the most capable systems from Anthropic and OpenAI, not just a budget alternative.
It natively supports text, image, and video input, offers a one million token context window, and was built specifically for software engineering, knowledge-intensive research, and long running agentic workflows. That combination of scale, context length, and multimodal support is a big part of why it is drawing so much attention.
The Company Behind Kimi K3: Moonshot AI
Moonshot AI is one of China’s most closely watched AI labs, and it is Moonshot’s most consequential release so far. The company first gained international notice in mid-2025 with Kimi K2, an open-weight coding model that undercut Western pricing while matching or beating some benchmarks from established players. Kimi K2 Thinking followed later that year with stronger agentic reasoning. K3 builds directly on that momentum, and the scale of the jump, from a well-regarded mid-size model to a 2.8 trillion parameter frontier system, is what has turned heads across the industry. Moonshot has said Kimi models have set the upper bound of open model sizes in nine of the past twelve months.
How Does Kimi K3 Actually Work?
This is the part most coverage skips, but it is what actually explains the benchmark jump. It is not simply a bigger version of an existing design. Moonshot built it around three architectural changes that work together: a sparser mixture of experts setup, a new attention mechanism, and a new way of passing information between layers.
A Sparse Mixture of Experts
Most large language models are dense, meaning every parameter gets used for every token they generate. Kimi K3 is a Mixture of Experts, or MoE, model instead. Rather than one giant network, it is built from 896 smaller specialized subnetworks called experts, and a router decides which experts should handle each token as it is processed. For this model, only 16 of those 896 experts activate for any single token, roughly 1.8 percent of the total. This is what lets the model carry 2.8 trillion total parameters while keeping the actual compute cost of generating each token far lower than a dense model of the same size would require.
Moonshot calls this routing system Stable LatentMoE, and it uses a technique called Quantile Balancing to decide which experts handle which tokens. Instead of relying on hand tuned rules to keep the workload spread evenly across experts, which can be unstable at this scale, Quantile Balancing derives the routing decisions directly from the router’s own score distribution. In plain terms, it is a more self correcting way of making sure all 896 experts get used efficiently instead of a handful becoming overloaded while others sit idle.
Kimi Delta Attention: Solving the Long Context Problem
The second innovation targets a well known weakness in how transformer models handle long text. Standard attention compares every token in a sequence against every other token, which becomes extremely expensive as context length grows. Processing a full one million token document with that approach is, in Moonshot’s own words, computationally brutal.
Kimi Delta Attention, or KDA, is a hybrid attention mechanism that mixes efficient linear attention with a smaller amount of full attention, reportedly in roughly a three to one ratio.
Linear attention is much cheaper to compute but historically loses some precision on tasks that need to recall specific details from far back in a long document. By blending the two approaches instead of choosing one, it gets most of the speed benefit of linear attention while keeping enough full attention to stay accurate.
Moonshot reports that KDA delivers up to 6.3 times faster decoding for million token contexts compared with standard attention, which is a major part of how the one million token context window stays usable in practice rather than being a number on a spec sheet nobody actually uses.
Attention Residuals: A More Efficient Way to Train
The third piece, Attention Residuals, addresses training rather than inference. In a standard transformer, information from earlier layers is passed forward to later layers through residual connections, added in a fairly uniform way at every layer. Attention Residuals instead let the model selectively retrieve representations from earlier in the network rather than accumulating them uniformly at every depth. Moonshot reports this yields roughly a 25 percent improvement in training efficiency for less than 2 percent in additional compute overhead, which matters enormously at the scale of a 2.8 trillion parameter model, where even small efficiency gains translate into meaningful savings in compute and time.
Together, Moonshot says these architectural changes produce roughly a 2.5 times improvement in overall scaling efficiency compared with Kimi K2, meaning the model converts training compute into usable capability more effectively than its predecessor did.
Reasoning Modes
K3 also runs as an always on reasoning model, meaning it works through problems step by step before answering rather than generating a response immediately. At launch, only a maximum reasoning effort setting was available. Moonshot has since documented three reasoning effort tiers, Standard, High, and Max, giving developers a way to trade off response speed against reasoning depth depending on the task.
Why This Combination Matters in Practice
None of these three changes would mean much on their own. A sparse mixture of experts setup alone just makes a huge model cheaper to run, but it does nothing for long documents. A faster attention mechanism alone helps with context length, but it does not make training a multi-trillion parameter network affordable. Put together, the three pieces solve different parts of the same underlying problem, which is how to build a model large enough to reach frontier level intelligence without making it too slow to use or too expensive to train.
This is also why independent reporting keeps circling back to the same three terms, Stable LatentMoE, Kimi Delta Attention, and Attention Residuals, whenever it tries to explain how a relatively young lab caught up to years of investment from much larger US companies so quickly. The architecture, not just the raw parameter count, is doing a meaningful share of the work. Moonshot has said a full technical report covering training details and further evaluation results will follow the public weights release.
What Else Makes Kimi K3 Different
Beyond the architecture, a few other choices set Kimi K3 apart from earlier Chinese releases and from many Western models as well.
Scale. At 2.8 trillion parameters, it is the largest open weight model publicly announced so far.
Long context. It ships with a 1,048,576 token context window, large enough to hold entire codebases, lengthy research documents, or extended multi-turn agent sessions without truncation.
Multimodal input. It accepts text, images, and video, which extends its usefulness beyond pure text generation into document analysis and visual reasoning tasks.
Open weights. Unlike the premium proprietary models it competes with, Moonshot has said the full Kimi K3 weights will be released publicly by July 27, 2026, following the open-source approach that has become a hallmark of Chinese AI development over the past two years.
Kimi K3 Benchmarks: How It Compares to Claude and GPT
The benchmark results are the main reason Kimi K3 became a story overnight. Within hours of release, it topped the Arena leaderboard for AI coding tools, becoming the first Chinese model to reach the number one spot on that influential platform. On Artificial Analysis, an independent benchmarking service, it ranked fourth overall and took the top spot on Design Arena.
According to reporting on the launch, Kimi K3 beat Anthropic’s Opus 4.8 and OpenAI’s GPT 5.5 on several benchmarks, including coding and general agent tasks. It still trails the newest flagship models from both companies on overall performance, but the gap has narrowed considerably compared to previous Chinese releases. Independent evaluation also found it strong on browsing and long document benchmarks, though it lagged behind GPT 5.6 Sol on a few reasoning heavy tests, including Terminal-Bench and GPQA Diamond.
The bigger story is not any single benchmark. It is how quickly it closed a gap that, a year earlier, looked much wider. Industry analysts have pointed to this as evidence that Chinese labs are no longer simply following behind, but are now shipping models that force US labs to respond.
Kimi K3 Pricing: What It Costs to Use
Unlike Kimi K2, which was known for aggressive discount pricing, it arrived priced like a premium model. The official API charges $3.00 per million input tokens and $15.00 per million output tokens, matching the sticker price of a leading US mid-tier model. Cached input tokens drop to $0.30 per million, a 90 percent discount that rewards workloads with stable system prompts or repeated document context, such as coding agents and long research sessions.
Pricing is flat across the full context window, so there is no premium tier once a prompt grows past a certain length, something that sets it apart from providers that charge more for very long inputs. Consumer access through the Kimi app remains free at standard rate limits, with paid membership tiers available for heavier use.
Compared with other Chinese open models, Kimi K3 sits at the expensive end. DeepSeek and GLM based models remain considerably cheaper. That pricing decision signals that Moonshot is positioning the model as a capability play rather than a race to the bottom on cost.
Show Image Alt text: Kimi K3 API pricing compared with Claude and GPT model pricing
Why Kimi K3 Matters for the Global AI Race
Kimi K3 lands at a moment when open weight Chinese models are already gaining serious ground internationally. Reports indicate that Chinese models recently overtook their US counterparts in token traffic on OpenRouter, a widely used AI model access platform, with Chinese models accounting for a majority share of related traffic in early July. Venture capital investors have also noted that a large share of companies adopting open source AI now choose Chinese models over Western alternatives.
For businesses and independent developers, the arrival of Kimi K3 adds another credible, openly available option alongside models from Anthropic, OpenAI, and Google. For policymakers, it reopens an ongoing debate in Washington about whether US companies should be building on or competing against Chinese open weight systems. Whatever position one takes in that debate, it is further proof that the gap between US and Chinese frontier AI has narrowed faster than many expected just a year ago.
Should You Try Kimi K3?
Whether Kimi K3 is worth adopting depends heavily on the use case. For teams already running coding agents or long document workflows, its context window and coding benchmark results make it worth testing, particularly through the cached pricing tier. For simpler chat or lightweight tasks, its pricing is no longer the bargain that earlier Kimi releases offered, so cheaper open models may be a better fit.
Developers curious about trying it can access Kimi K3 through the Kimi web app, the Kimi mobile apps, or the official API, with open weights expected to follow shortly after launch. Anyone comparing frontier options may also want to revisit how it fits alongside the broader landscape of large language models before committing a production workload to it.
Moonshot’s own Kimi K3 tech blog cited where the architecture is explained
The official API docs – cited in the pricing section
MarkTechPost’s launch coverage – cited in the benchmarks section
Explore more from Blogs AI Learning Hub Subscribe to AI Pathway Lab for more guides, tutorials and news latest
Frequently Asked Questions
What is Kimi K3?
Kimi K3 is a 2.8 trillion parameter open weight AI model released by Moonshot AI on July 17, 2026. It uses a sparse mixture of experts architecture, supports text, image, and video input, and offers a one million token context window.
Who made Kimi K3?
Moonshot AI, a Beijing based startup backed by Alibaba, built Kimi K3. Moonshot previously released Kimi K2 and Kimi K2 Thinking, both known for strong coding performance at low cost.
How does Kimi K3 work?
Kimi K3 combines a sparse mixture of experts design that activates only 16 of 896 experts per token, a hybrid attention mechanism called Kimi Delta Attention for efficient long context processing, and a technique called Attention Residuals that improves training efficiency across the network’s depth.
Why is Kimi K3 able to handle a one million token context window?
Kimi Delta Attention makes this practical by blending fast linear attention with a smaller share of full attention, which Moonshot reports delivers up to 6.3 times faster decoding at long context lengths than standard attention alone.
Is Kimi K3 free to use?
It is free to try through the Kimi app and website at standard rate limits. Heavier use through the API is billed per token, and consumer subscription tiers are available for higher limits.
How does Kimi K3 compare to Claude and GPT?
Kimi K3 has outperformed some benchmarks from Claude Opus 4.8 and GPT 5.5, particularly in coding and agent tasks, but it still trails the newest flagship models from Anthropic and OpenAI on overall performance.
Is Kimi K3 open source?
It was released as an open weight model, meaning its parameters are available for developers to download and customize, with full weights expected by July 27, 2026.
How much does Kimi K3 cost through the API?
Kimi K3 costs $3.00 per million input tokens and $15.00 per million output tokens, with a discounted $0.30 per million rate for cached input, applied flatly across its full context window
Want to understand how models like Kimi K3 fit into the bigger picture? Check out our guide to the future of AI: 15 trends and predictions, our breakdown of ChatGPT vs Claude vs Gemini vs Copilot, and our list of free AI courses with certificates. If you are new to how these models actually work, start with our prompt engineering guide and our overview of AI chatbots.
Sources: CGTN, CNBC, Bloomberg, Axios, Artificial Analysis, Kimi tech blog