BEIJING Chinese AI lab Moonshot AI announced Kimi K3 this week, a 2.8 trillion parameter model that immediately claims the title of the largest openly available AI model ever released. The company is calling it the “first open 3T-class model,” rounding up from 2.8 trillion, and the benchmarks suggest they have a legitimate claim to frontier status.
Why it matters: The open-weight AI arms race just escalated dramatically. Kimi K3 doesn't just compete with frontier models from OpenAI and Anthropic it beats them on several key benchmarks, including frontend code generation. And at a time when the narrative has been that Chinese AI labs compete on price alone, Kimi K3 arrives with premium pricing that challenges that assumption.
The model is currently available through Moonshot's website and API at $3 per million input tokens and $15 per million output tokens. That pricing puts it at parity with Anthropic's Claude Sonnet series, making it the most expensive model ever released by a Chinese AI lab. It's a significant departure from their previous models Kimi K2.6 was priced at $0.95/$4 per million tokens, roughly one-third the cost.
The numbers are staggering.
Kimi K3's 2.8 trillion parameters more than double Kimi K2.6's 1 trillion and surpass DeepSeek's 1.6-trillion-parameter V4 Pro, which previously held the open-weight size crown. Moonshot claims the model was trained on a cluster of over 100,000 H100-equivalent GPUs, a training run that would have cost well over $1 billion at current compute prices.
Simon Willison, the developer and blogger who has tracked model performance through his informal “pelican benchmark” for nearly two years, tested Kimi K3 and noted its unusual inference characteristics. “It only has one reasoning effort right now, 'max' and it shows,” Willison wrote. “The model consumed 13,241 reasoning tokens to output 3,417 tokens of response. That pelican cost 25 cents!” The high reasoning overhead means that while Kimi K3's raw capability is impressive, it comes at a computational cost that may limit its use cases.
Benchmark performance that demands attention.
Moonshot's self-reported benchmarks show Kimi K3 mostly beating Claude Opus 4.8 max and GPT-5.5 high across standard evaluation categories, while losing to Claude Fable 5 and GPT-5.6 Sol at the very top of the frontier. The model is particularly strong in coding tasks it now holds the #1 position on Arena.ai's Frontend Code arena, surpassing even Claude Fable 5 for frontend development workloads.
Artificial Analysis, an independent model evaluation platform, tested Kimi K3 on their private long-horizon knowledge work evaluation. The results were striking: Kimi K3 achieved an overall Elo of 1547, a gain of +732 points from Kimi K2.6, trailing only Claude Fable 5 among tested models. In terms of cost per task ($0.94), it's similar to GPT-5.6 Sol ($1.04), roughly half the price of Opus 4.8 ($1.80), but higher than other open-weight peers.
Perhaps most notably, Kimi K3 is significantly more efficient with tokens despite its massive parameter count. Artificial Analysis found that the model uses 21% fewer output tokens than K2.6 on their evaluation, suggesting that the architectural improvements Moonshot made between versions are substantive, not just a scaling story.
The open-weight promise.
Moonshot has committed to releasing Kimi K3's weights under an open license by July 27, 2026 just over a week from now. If they follow through, it will be the largest open-weight model ever released by a wide margin. The open-weight release would allow developers to self-host the model, fine-tune it for specialized applications, and build on top of it without API dependency.
This commitment mirrors the strategy that has made DeepSeek and Meta's Llama series so influential. Open-weight releases create an ecosystem of fine-tuned variants, community tooling, and research that compounds the model's value far beyond what a closed API can deliver. On Hugging Face alone, Llama 5 already has over 180,000 fine-tuned variants, creating combinatorial innovation at a scale no single lab can match.
But Kimi K3's size creates practical challenges for the open-weight promise. A 2.8-trillion-parameter model requires approximately 5.6 terabytes of memory at FP16 precision. Running inference at acceptable speeds requires multiple high-end GPUs in parallel a configuration that most developers and small startups simply cannot afford. Quantization to INT4 would reduce the memory requirement to roughly 1.4 terabytes, but even that requires a cluster of A100 or H100 GPUs.
“The open-weight release matters more for the ecosystem than for individual developers,” said Dr. Maria Kowalski, AI researcher at Cambridge. “It means that labs with compute resources can fine-tune and distill this model. We'll see dozens of smaller, more efficient derivatives within months. The real value of open-weight frontier models is the downstream innovation they enable.”
What this means for the AI landscape.
Kimi K3's arrival reshapes the competitive dynamics of the AI industry in three critical ways.
First, it destroys the narrative that Chinese AI labs can only compete on price. At $15 per million output tokens, Kimi K3 is priced at a premium that puts it in direct competition with Anthropic and OpenAI's top-tier offerings. Moonshot is betting that benchmark performance, not cost, is what matters to serious AI developers and the early data suggests they may be right.
Second, it puts pressure on every AI lab to deliver their next-generation models faster. OpenAI and Anthropic have been the undisputed frontier leaders for two years. Kimi K3 demonstrates that the gap has narrowed to the point where a challenger can claim benchmark leadership in specific categories. The response from Western labs will likely be accelerated release schedules and aggressive pricing of their own.
Third, it validates the open-weight strategy at the highest tier of model scale. DeepSeek proved that open-weight models could be competitive at the 1-trillion-parameter scale. Kimi K3 extends that proof to nearly 3 trillion parameters. If open-weight models continue to match or exceed proprietary alternatives, the economic case for proprietary APIs weakens significantly.
The bottom line.
Kimi K3 is not a “cheap alternative.” It's a frontier model with premium pricing, legitimate benchmark leadership, and an open-weight commitment that could reshape the AI supply chain. For AI developers and founders, the calculus is becoming more complex: you can build on proprietary APIs from OpenAI or Anthropic, you can self-host open-weight models from Meta or DeepSeek, and now you have a third option from Moonshot that sits at the intersection of both strategies.
The model's true impact will depend on whether Moonshot delivers the open-weight release by July 27, and whether the developer community can overcome the practical challenges of running a 2.8-trillion-parameter model. But one thing is already clear: the era of a two-player frontier market is over.