Kimi K3 Open Weights Go Live — 2.8T Parameters, Free to Anyone Who Can Run Them
Kimi K3‘s open weights are live and available to download.
That makes Moonshot AI‘s flagship the largest open-weight AI system anyone has ever released — 2.8 trillion parameters, free to any developer or organization with the hardware to run it.
Also Read: CXMT Soars 500% in Shanghai Debut — Now the Mainland’s Most Valuable Listed Company
No license fee. No gatekeeping. Just the weights.
Moonshot published them on July 27, at a moment when the company is chasing a $50 billion valuation and weighing a Hong Kong IPO.
Kimi K3 Open Weights Rewrite The Scale Ceiling
This release is the first time a model at the 2.8-trillion-parameter scale has been made freely downloadable. To put that number in context, a parameter is a single learned numerical value inside a neural network.
During training, a model adjusts billions or trillions of these values to capture patterns in text, code, and data. More parameters generally mean a model can store and reason over more complex relationships, though the relationship is not perfectly linear.
For comparison, Meta‘s Llama 4 Maverick, the most capable openly available Western model before this release, tops out at around 400 billion active parameters in its mixture-of-experts configuration.
Kimi K3 uses a similar mixture-of-experts architecture, where only a subset of parameters activate for any given input, keeping inference costs manageable despite the enormous total count. Even so, running K3 locally demands significant GPU memory, and most developers will access it through API endpoints rather than self-hosting.
A Bloomberg report published July 27 confirmed the release, citing Moonshot’s intent to expand its reach in the global open software community.
Why Silicon Valley Is Treating Kimi K3 Open Weights As A Threat
The reaction inside US AI labs has been swift.
Developers and commentators speaking on the t3.gg podcast said Anthropic and OpenAI are closely watching this release, with one analysis framing K3 as evidence that Chinese labs are closing the capability gap at the open-weights tier faster than US counterparts anticipated.
The strategic logic is straightforward. Closed frontier models like GPT-4o and Claude 4 Opus command premium pricing because they sit at the capability frontier and are not downloadable.
When an open-weight model approaches that frontier, it pressures the business model of every closed-API provider. Developers who can run a comparable model on their own infrastructure have no reason to pay per-token rates to an American lab.
Moonshot is targeting a $50 billion valuation, according to a Whalesbook analysis published July 27.
Analysts cited by Whalesbook said Moonshot is targeting a Hong Kong IPO as its next major financial milestone. Releasing Kimi K3 open weights builds the developer ecosystem that makes that valuation credible.
How Moonshot Built A Model This Large
Mixture-of-experts, the architecture behind K3, works differently from a dense transformer.
A dense model activates all of its parameters for every token it processes. A mixture-of-experts model routes each token through a learned gating function that selects only a small group of specialist subnetworks, called experts, to handle that input.
The result is a model with a very large total parameter count but a much smaller active parameter count at inference time.
This matters economically. A 2.8-trillion-parameter dense model would be economically impossible to serve at API prices competitive with Western labs.
By routing tokens through a fraction of the total network, Moonshot can offer K3 at costs close to smaller dense models while retaining the capacity headroom that comes with a larger expert pool.
The Stratechery Sharp Tech podcast, recorded July 23, devoted a full segment to K3 and the future of open-weights AI in the US market. The hosts said the US open-source AI market has stagnated partly because major US labs have moved toward closed proprietary releases, leaving a gap that Chinese labs are filling.
The Road From Moonshot Lab To Global Download
Moonshot AI was founded in 2023 and launched the Kimi chatbot as its consumer product in China.
The company grew quickly on the strength of Kimi’s long-context capabilities, which allowed it to process documents far longer than most competing models at launch. By mid-2025, Kimi had tens of millions of monthly active users in China.
The decision to release Kimi K3 open weights globally represents a strategic pivot.
Rather than keeping its most capable model proprietary, Moonshot is betting that a massive open-weights release will drive developer adoption, generate benchmark citations, and build the brand recognition needed to support a Hong Kong listing at a $50 billion valuation. That playbook mirrors what Meta did with Llama, which turned an initially skeptical reception into a dominant position in enterprise AI infrastructure by releasing weights freely and letting the developer community build on top.
The key difference is regulatory context.
US lawmakers have grown increasingly concerned about Chinese AI models circulating freely in American infrastructure. A model of K3’s scale released from a Chinese lab will draw scrutiny from the Commerce Department and potentially trigger export-control debates in reverse, focusing on whether US companies should be permitted to build on top of Chinese open-weight models in sensitive applications.
What Kimi K3 Open Weights Mean For The Compute Market
The release also carries implications for GPU compute demand.
Running the model at scale requires high-end accelerator clusters. Every organization that downloads and self-hosts K3 rather than using a closed API is consuming GPU capacity on its own hardware or through a cloud provider.
That redirects revenue from API-based AI labs toward GPU cloud and data center operators.
For the open-weights ecosystem broadly, this release sets a new benchmark for what the tier can deliver. Developers who have treated open-weight models as a cost-saving measure rather than a genuine capability tool may revise that view.
A 2.8-trillion-parameter model that performs competitively with closed frontier systems changes the build-versus-buy calculus for every enterprise AI team.
Read Next: OpenAI CEO Says the Singularity Is Here, and He Is Not Speaking Hypothetically
