Moonshot AI Just Released the Largest Open-Weight Model Ever Built
Moonshot AI published the full weights of Kimi K3 on July 27, putting all 2.8 trillion parameters of its flagship model into public hands under a Modified MIT license. The Beijing-based lab had been running K3 through its API since July 16, and the weight release makes it, by Moonshot's own description in its technical blog, the world's first open 3-trillion-class system and the largest open-weight AI model released to date.
K3 is a mixture-of-experts model, and the 2.8 trillion figure describes capacity, not the compute it burns per answer. The model activates only 16 of its 896 experts on any given token, roughly 50 billion parameters of live compute per step, paired with an attention scheme Moonshot calls Kimi Delta Attention and a second technique it labels Attention Residuals. The company says the combination delivers about a 2.5 times improvement in scaling efficiency over its previous flagship, Kimi K2, and the model carries a 1-million-token context window with native vision support.
On benchmarks, K3 ranked second overall on the Vals AI index and third on Artificial Analysis's Intelligence Index, trailing only Claude Fable 5 and GPT-5.6 Sol Max while undercutting both on price, according to figures AI researcher Nathan Lambert compiled at Interconnects. It took first place outright in the LMArena Frontend Code evaluation, scoring 1,679 points and jumping 17 spots past its predecessor, Kimi K2.6, in blind developer testing.
Open weights, closed hardware
The number drawing the most scrutiny is not on the benchmark card. Even compressed into MXFP4, a four-bit format, the full weight set needs roughly 1.4 terabytes of fast memory before the model reads a single token of a prompt, according to analysis published by TECHi. In 16-bit precision the same weights would run closer to 5.6 terabytes. That points to something on the order of eighteen 80-gigabyte accelerators just to hold the model resident, which rules out a workstation and points instead to racks of Nvidia Blackwell or AMD MI400 silicon, the two platforms MXFP4 runs on natively.
That gap between "downloadable" and "runnable" is the practical story of this release. Most organizations that adopt K3 will rent it through Moonshot's API or a cloud provider hosting the weights on their behalf, inheriting a version of the same vendor dependency that open weights are supposed to remove, just with more suppliers to shop between.
An export-control backdrop
Moonshot's own disclosures add a layer Tom's Hardware flagged directly: the company ran its kernel optimization benchmarks on Nvidia's H200 and an unnamed "GPGPU from an alternative vendor," and benchmarked its from-scratch compiler, MiniTriton, against Nvidia's Triton on an L20, the cut-down Ada-generation card sold into China under U.S. export rules. The blog does not say where the H200 hardware is physically located, a detail that matters after Congress passed a bill in January closing the offshore cloud rental loophole that had let Chinese firms remotely access restricted accelerators.
The release also lands as U.S. policymakers weigh whether to restrict Chinese open-weight models altogether. Axios reported on July 20 that the Commerce Department had considered adding multiple Chinese AI labs to its Entity List, which would cut off U.S. access without a license, and that the White House had floated an executive order requiring U.S. companies to guarantee security and accept liability before hosting Chinese models. None of those measures have been enacted.
Anthropic accused Moonshot in February of using 3.4 million Claude exchanges to train its models through distillation. K3 now benchmarks within a few points of the systems named in that complaint, a result Lambert argues cuts against the idea that Moonshot's gains rest mainly on copying rather than original scaling and architecture work.
Independent labs have not yet reproduced Moonshot's benchmark and efficiency claims on hardware outside the company's control, and the license text published alongside the weights is the first point at which businesses can evaluate exactly what commercial use it permits.
Sources: Tom's Hardware · TECHi · Interconnects AI · Axios · Moonshot AI