Moonshot AI released the full weights of Kimi K3 at 00:00 UTC on July 27, 2026, making its 2.8-trillion-parameter model the largest open-weight AI system ever made freely downloadable.
The model had been available only through Moonshot’s API since July 16; the weights release marks the point where any organization can, in theory, host the model rather than pay per token.
In theory is doing a lot of work in that sentence. The full weights run to roughly 1.4 terabytes even in Moonshot’s compressed MXFP4 four-bit format — about 5.6 terabytes at standard 16-bit precision — and because Kimi K3 is a sparse mixture-of-experts model with 896 experts, all of them have to sit in fast memory at once regardless of how few are active per token, as one technical breakdown explained.
Moonshot recommends a configuration of 64 or more accelerators on a single high-bandwidth interconnect, hardware that natively supports MXFP4 only on Nvidia’s Blackwell generation or AMD’s MI400 chips.
Free to download, expensive to use
That leaves self-hosting realistically limited to cloud providers, large research labs, and well-funded enterprises already running serious inference infrastructure — not hobbyists or small startups.
For that tier of buyer, though, the calculation genuinely shifts: instead of paying Moonshot, OpenAI, or Anthropic per token indefinitely, a company can pay once for hardware and engineering and then serve unlimited queries on its own infrastructure. Open weights don’t make inference free; they make the margin visible, and shift it from the model vendor to whoever owns the hardware.
The context nobody should skip
Kimi K3’s rise has not been free of controversy. White House officials have publicly alleged that Moonshot accessed export-restricted Nvidia chips through Thailand-based infrastructure and distilled Anthropic’s model during training — allegations Moonshot has not directly addressed, and which several independent AI researchers have publicly disputed.
As of July 25, no formal enforcement action against Moonshot had been announced, and the weight release proceeded on its original schedule regardless. Independent testers have also flagged a hallucination rate around 51% on certain evaluations that didn’t appear in Moonshot’s own published benchmark charts — a reminder that leaderboard position and reliability in production are not the same thing.
What to watch next
- Whether major cloud providers move quickly to offer hosted Kimi K3 inference, closing the gap for companies that can’t self-host.
- How enterprises weigh the data-sovereignty question, given China’s National Intelligence Law applies to Moonshot’s hosted API even if the open weights themselves don’t.
- Whether independent benchmarks converge with or continue to diverge from Moonshot’s own published results.
Helpful Sources
- Kimi K3’s open weights arrive July 27. The catch is 1.4TB — TECHi
- Moonshot AI releases Kimi K3 open weights, world’s largest open-weight model — Startup Fortune
- Kimi K3 Open Weights Drop July 27: Near-Frontier Coding, Undisclosed Hallucination Risk — Tech Times
Disclaimer: This content is meant to inform and should not be considered financial advice. The views expressed in this article may include the author’s personal opinions and do not represent Times Tabloid’s opinion. Readers are advised to conduct thorough research before making any investment decisions. Any action taken by the reader is strictly at their own risk. Times Tabloid is not responsible for any financial losses.
Follow us on X, Facebook, Telegram, and Google News
