Kimi K3: Has Chinese AI Finally Caught Up With America?

On July 16th, 2026, Moonshot released Kimi K3, an open-source 2.8T Mixture-of-Experts (MOE) model with a context window of up to 1 million tokens. It is the largest open-weights model to ever be released, and also the smartest. It takes the crown from DeepSeek's 1.6T v4 Pro, which held the title for months. But can it actually compete with American AI models, or is this just another case of a Chinese lab putting up impressive numbers on paper?

Now let's talk about the architectural breakthroughs the team at Moonshot made while training this model, because the engineering here is legitimately interesting.

The first big idea is called Kimi Delta Attention, or KDA. Think of it as giving the model a smarter memory. Instead of trying to remember everything equally, it learns what to keep and what to let go of as it reads through huge amounts of text. KDA is a hybrid linear attention mechanism that Moonshot states enables up to 6.3x faster decoding in million-token contexts. That is how K3 can handle up to a million tokens of context without slowing down or losing track of details. For context, most frontier models still struggle with truly long inputs because standard attention mechanisms scale quadratically with sequence length. KDA gets around this by being selective about what information it holds onto, which is a fundamentally different approach to long-context processing. Because KDA poses new challenges for prefix caching, Moonshot contributed an implementation to vLLM, which shows they are thinking about practical deployment, not just benchmark scores.

The second idea is called Attention Residuals, or AttnRes. Normal AI models are built in layers, and information can get watered down as it passes through so many layers, kind of like a message getting fuzzy after being repeated too many times. AttnRes works along the depth axis, selectively retrieving representations across depth rather than accumulating them uniformly. This lets each layer look back and pull out exactly what it needs from earlier layers instead of just blending everything together, so important details do not get lost as the model gets deeper. And the efficiency gains are real. Moonshot states AttnRes delivers roughly 25% higher training efficiency at under 2% additional cost. Getting 25% more training efficiency for almost no extra compute is the kind of innovation that actually changes the economics of building these models.

The third idea is about efficiency and sparsity. The model has an enormous number of parameters, but it does not use them all at once. Sparsity is the third lever, with K3 using Stable LatentMoE and effectively activating only 16 of 896 experts. For every piece of text it processes, it only wakes up this small handful of specialized experts instead of the whole system. This keeps the model fast and affordable to run even though it is massive. Quantile Balancing derives expert allocation directly from router-score quantiles, which eliminates heuristic updates and a sensitive balancing hyperparameter. If you are not deep into ML research, the takeaway is simple: they figured out a better way to decide which experts to turn on for each task, and that makes the whole system more stable during training. Together with refined training and data recipes, these changes yield roughly 2.5x better overall scaling efficiency than Kimi K2.

These three innovations, KDA, AttnRes, and Stable LatentMoE, are what separate K3 from just being a bigger version of K2. They actually rethought how information flows through the model both in terms of sequence length and depth, which is a big part of why it performs so well on tasks like coding and multi-step reasoning.

So how good is the model, actually? Kimi K3 achieves a score of 57 on the Artificial Analysis Intelligence Index, a composite benchmark that evaluates models across reasoning, knowledge, mathematics, and coding. That is a landmark number. It only gets beaten by Claude Fable 5 at 60 and ChatGPT Sol at 59. The model outperformed Claude Opus 4.8 max and GPT-5.5 high on benchmarks, which are models that, until very recently, were considered the state of the art.

But it is not just about aggregate scores. K3 has specific areas where it genuinely leads the pack. Kimi K3 topped the Frontend Code Arena with a 76% pairwise win rate, meaning when its output was compared head-to-head against other models on the same task, it was picked as the better output 76% of the time on average. For reference, Claude Fable 5 hit 63% and GPT-5.6 Sol hit 58%. Arena's announcement puts K3 at number 1 on its Frontend Code Arena, a 17-place jump from Kimi K2.6's number 18, past Claude Fable 5. It topped six out of seven sub-domains assessed, including Brand and Marketing and Data and Analytics.

Beyond frontend coding, the model also put up strong numbers on agentic benchmarks. At launch it scored 93.5% on GPQA Diamond, the strongest open-weight result on that benchmark published at the time, alongside 88.3% on Terminal-Bench 2.1. On agentic work, it hit 91.2% on BrowseComp, the best published score on the tracker at release. In tests of real-world task automation, Kimi K3 ranked first in four out of eight benchmarks, including Automation Bench, SpreadsheetBench 2, and BrowseComp, while finishing second to Fable 5 in most others.

However, the fact that K3 is so intelligent comes with a price. Pricing for Kimi K3 is $3.00 per 1M input tokens and $15.00 per 1M output tokens. That makes it the most expensive Chinese AI model by a long shot. K3 costs $15 per million output tokens, compared to $4.40 per million output tokens for z.ai's GLM-5.2 and $0.87 for DeepSeek V4. So within the Chinese ecosystem, K3 is playing in a completely different price bracket.

Nonetheless, when compared to American AI models, the price makes a lot more sense. Fable 5 costs $10/M input tokens and $50/M output tokens, and 5.6 Sol costs $5/M input tokens and $30/M output tokens. Because of the similar intelligence but lower price, K3 cost only $0.95 per task on Artificial Analysis's evaluation, compared to Fable 5's $2.75 and 5.6 Sol's $1.04. If users hit cache memory, the pricing drops to thirty cents per million input tokens, which makes repeated interactions with the same context extremely cheap. For companies doing a lot of retrieval-augmented generation or running agents that revisit the same codebases, that cache pricing could be a game changer.

Moonshot AI has confirmed that Kimi K3 will be an open-weight model, with the full 2.8 trillion parameter model weights officially releasing on July 27, 2026. Once the weights are out, independent researchers can run their own evaluations and we will get a much clearer picture of where K3 truly stands.

Moonshot's previous AI models were already making inroads into Silicon Valley. Cursor, the vibe-coding startup, used Kimi to help build Composer, its AI coding agent. DoorDash also delegates lower-level work to Kimi K2.6. Thinking Machines tapped Kimi K2.5 to generate early post-training data for its new Inkling model. So American companies are already using Moonshot's models in production, and K3 is only going to accelerate that trend.

The release was timed to land just ahead of the 2026 World Artificial Intelligence Conference in Shanghai, which is not a coincidence. Moonshot is making a statement, not just to the AI research community but to the broader geopolitical conversation about who is winning the AI race.

Overall, Kimi K3 is probably the most important Chinese model release of this year so far. It doesn’t just compete with American frontier models on benchmarks. It beats them in specific areas like frontend coding. It is meaningfully cheaper to run. And once the weights drop on July 27th, anyone in the world can download it and build on top of it. The gap between open-source Chinese models and closed-source American ones has never been smaller, and that has massive implications for everyone building with AI right now.