The kimi k3 open weights dropped on July 26, a day ahead of schedule, making Moonshot AI the provider of the largest publicly available open-weight model. Developers can now download, modify, and self-host a 2.8 trillion-parameter mixture of experts model with a 1-million-token context window. The Beijing startup released the weights through its Hugging Face organization alongside a detailed technical report, confirming the raw scale the company had teased a week earlier.
Kimi K3 Open Weights Release
Moonshot originally promised the weights by July 27. They landed at roughly 7:30 p.m. EDT on July 26, beating the target by a day. The core model activates only 104 billion of its 2.8 trillion parameters per token, a result of a sparsity design that fires 16 of 896 experts in its MoE architecture. This makes Kimi K3 not just a headline figure. It is the first open model to cross the 3 trillion-parameter class, and its 2.5× scaling efficiency gain over the previous K2 model means more intelligence per compute cycle.
The full package includes native vision and that million-token context window. Researchers and tinkerers can feed entire codebases or lengthy legal documents into a single prompt. Moonshot’s technical report walks through Kimi Delta Attention and Attention Residuals, two architectural upgrades that improve information flow across long sequences and deep model layers. The weights ship in MXFP4 weight format with MXFP8 activations, a practical concession to storage and memory constraints.
Benchmarks Outrun Older Frontier Models
Moonshot’s own benchmarks and independently tracked arena results place Kimi K3 behind Claude Fable 5 and GPT-5.6 Sol on overall intelligence but ahead of every other open or closed model tested. On the DeepSWE coding task, for instance, Kimi K3 scored 67.5, topping Claude Opus 4.8’s 59.8 and trailing GPT-5.6 Sol by only 5.5 points. Terminal-Bench 2.1 showed a similar pattern: Kimi K3 hit 88.3, jumping past Claude Fable 5’s 78.0 and roughly matching Sol at 88.8.
Long-horizon coding sessions are where the kimi k3 open weights shine in practice. Moonshot demonstrated the model writing a Triton-like compiler called MiniTriton from scratch, complete with a tile-level IR layer and PTX code generation. Kernel optimization tasks that stretched over 24 hours put Kimi K3 ahead of GPT-5.5 and competitive with Fable 5 when the latter had fallback enabled. On SWE-Marathon, which tests sustained software engineering work, Kimi K3 scored 42.0, edging out Sol’s 39.0 though still behind Opus 4.8’s 68.0.
Agent and vision tasks also show consistent gains. Kimi K3 reached 91.2 on BrowseComp and 30.8 on AutomationBench, both numbers exceeding Claude Fable 5 and the other reference models. These results matter because they come from a model that anyone can download and run locally. Developers who want frontier-like performance without a closed API now have a new baseline to measure against.
Custom License Replaces MIT Style
The earlier Kimi K2 shipped with what Moonshot called a “Modified MIT” license. The new weights arrive with a custom document that gates commercial inference use. If a company’s annual revenue exceeds $20 million, the default terms do not cover hosting the model for production workloads. Smaller startups, academics, and hobbyists remain unaffected, but any business scaling quickly will need to negotiate with Moonshot or use the hosted API. Kimi K3’s release lands the same week as Google’s open-weight Gemma 4 models, though Gemma ships under a fully permissive Apache 2.0 license with no revenue threshold.
The shift went largely unnoticed in the initial rush of download links and benchmark charts. The license itself sits in the Hugging Face repository as a standalone file. It is not hidden, but early social media chatter assumed the MIT spirit would carry over. Moonshot’s choice mirrors a broader conversation about what “open” means when trillion-parameter models cost millions to train. The kimi k3 open weights license attempts to balance adoption with a path to monetization.
The license threshold puts Moonshot in a different category from fully permissive models like Llama or Mistral. Developers evaluating Kimi K3 for commercial deployment will need to check their revenue run rate first. The hosted API, which does not carry the same restriction, already powers the Kimi chatbot and Kimi Code product, where Moonshot says daily revenue has grown by a factor of at least six since K3 launched.
Moonshot Revenue Spikes and IPO Plans
Moonshot reached $300 million in annual recurring revenue in June, up from $200 million in April. The jump lines up with the initial hosted rollout of Kimi K3 on July 16. That open-weight release a week later adds a self-serve distribution channel that the company expects will widen its developer base further. Founder Yang Zhilin told Bloomberg that openness is the strategy to grow users faster than U.S. rivals who keep frontier models behind APIs.
Investors are paying attention. Moonshot is reportedly seeking fresh funding at a $50 billion valuation, more than double its previous private valuation. A Hong Kong initial public offering could happen as soon as this year. The commercial restrictions in that license suggest Moonshot wants the free distribution to feed a funnel that eventually converts large-scale users into paying API customers.
The raw numbers are striking for a model company that shipped its first public model only a year and a half ago. The kimi k3 open weights put the company’s technology in the hands of thousands of developers overnight while keeping a direct line to enterprise revenue through the commercial license trigger. That dual track is becoming a template for Chinese AI labs aiming for global reach without giving away the store.
White House Scrutiny Shadows Launch
White House Office of Science and Technology Policy Director Michael Kratsios accused Moonshot last week of leveraging American technology, including Nvidia hardware subject to export controls, to train the model. The statement added a geopolitical echo to the Kimi K3 launch. Moonshot has not detailed its compute supply chain, and the technical report does not mention chip origin.
The timing highlights a tension that public weight releases often skirt. Open distribution means a model can be used anywhere, but the underlying training might rely on infrastructure that falls under trade restrictions. Moonshot’s decision to push the release forward without addressing those questions leaves the model in a grey zone for some U.S. enterprise buyers.
Developers who downloaded the weights over the weekend are unlikely to worry about trade law when spinning up a coding assistant. But the scrutiny will intensify if the model becomes the default open-weight choice for long-context work. Moonshot’s $50 billion valuation target and potential public listing mean the political dimension is now part of the product story.
FAQ
Does Kimi K3 outperform GPT and Claude?
Kimi K3 trails Claude Fable 5 and GPT-5.6 Sol on overall intelligence benchmarks but beats every other tested model, including Claude Opus 4.8 and GPT-5.5, on many coding and agent tasks. The kimi k3 open weights give developers a model that sits between the prior generation of frontier models and today’s absolute best proprietary systems.
What are the hardware requirements to run Kimi K3?
Kimi K3 uses MXFP4 precision and activates 104B parameters per token. That demands substantial GPU memory; a single inference node typically needs several high-memory GPUs. Moonshot has not published minimum specs, but community inference partners are already working on optimized runtimes. Expect hardware requirements similar to other 100B-plus active parameter models.
Can I use Kimi K3 commercially?
Yes, with a catch. The license permits commercial use up to $20 million in annual company revenue. Beyond that, you must either license the model separately or use the paid Kimi API. The commercial threshold is the biggest departure from the “Modified MIT” license that shipped with Kimi K2.
