Moonshot AI drops open weights for Kimi K3, the largest downloadable AI model ever at 2.8 trillion parameters

Moonshot AI will release the full open weights of Kimi K3 on July 27, making the 2.8-trillion-parameter model the largest openly downloadable AI system ever shipped 1. The headline reads as a win for open AI. The download is 1.4 terabytes.

That gap is where this release actually lands for anyone deciding what to build on. Kimi K3 has been answering API calls since July 16 2. The weights that would let independent researchers verify Moonshot's benchmark claims arrive eleven days later 1. When they do, the binding question is not whether you can download them. It is whether you can keep them resident in fast memory long enough to serve a single request.

What 2.8 trillion parameters costs to run

Kimi K3 is a mixture-of-experts model. Only 16 of its 896 experts fire on any given token, which works out to roughly 50 billion parameters of live computation per step out of a 2.8-trillion-parameter total 3. The trillions describe how much the model knows. They do not describe how much math it does to answer you.

Moonshot pairs that sparse routing with two architectural changes it calls Kimi Delta Attention, a hybrid linear-attention scheme, and Attention Residuals, which alters how information passes between layers 3. The company also claims roughly a 2.5-times improvement in scaling efficiency over Kimi K2 4.

This split between total and active parameters produces two different cost profiles. Per-token inference looks like a mid-size model because only a fraction of the network activates. Hosting looks like something far larger, because all 2.8 trillion parameters must stay loaded in memory regardless of which experts are currently working.

The 1.4-terabyte gate

The compressed weights occupy about 1.4 terabytes, and they only reach that figure because Moonshot ships them in MXFP4, a four-bit floating-point format 4. At sixteen-bit precision the same weights would need roughly 5.6 terabytes 4. Even at four bits, the full set has to live in fast memory before the model reads a single token. The one-million-token context window adds to that footprint from there 3.

MXFP4 runs natively on Nvidia's Blackwell generation and AMD's MI400 accelerators 4. The hardware floor for running these weights locally is therefore the newest data-center silicon, not a workstation or a single GPU. Holding 1.4 terabytes implies roughly eighteen 80-gigabyte accelerators just to load the model, before any memory is set aside for context or concurrent users 4. A single rack of the newest cards, eight at 192 gigabytes each, can barely contain the weights with almost nothing left over 4. Moonshot recommends serving K3 on supernodes of 64 or more accelerators 3.

Who benefits, and who rents

The workflow that made smaller open models genuinely portable, running weights on a local machine, does not extend to 1.4 terabytes. The organizations positioned to self-host Kimi K3 are inference providers, large enterprises, and cloud operators who already maintain racks of Blackwell or MI400 silicon 4.

For a buyer weighing open against closed, that narrows the decision. The appeal of open weights is straightforward: no per-token fee set by one vendor, no capability pulled from a shipping product, the ability to run inference inside your own security perimeter. In practice, most teams that adopt Kimi K3 will pay per token through Moonshot's API or through a hosting cloud 4. The freedom to move between suppliers is genuine. It stops short of the self-sufficiency that makes smaller open models disruptive.

Moonshot's API pricing for K3 runs $0.30 per million cache-hit input tokens, $3 per million on cache misses, and $15 per million output tokens 3. That puts K3 in the same price band as Anthropic's Claude Sonnet series and makes it the most expensive model a Chinese AI lab has shipped to date 2.

Benchmark claims, now verifiable

Moonshot's own evaluation shows K3 outperforming Claude Opus 4.8 and GPT-5.5 on coding and agentic tasks, while still trailing Claude Fable 5 and GPT-5.6 Sol overall 3. The model claimed first place in Arena's Frontend Code evaluation at 1,679 points, ahead of Fable 5 in blind developer testing 3. On the Artificial Analysis intelligence index, K3 reached an Elo of 1,547, a 732-point jump from Kimi K2.6 2.

Once the weights are public on July 27, independent labs can test those numbers directly instead of taking Moonshot's word. Until then, they remain vendor-reported. The license terms that will determine whether commercial use is permitted are also still unpublished 4.

What frontier open weights signal now

Kimi K3 takes the size record from DeepSeek's 1.6-trillion-parameter v4 Pro 2. Bank of America analysts led by Alex Liu argue that large-scale pre-training combined with architectural innovation can still deliver step-change gains for Chinese labs operating under U.S. compute restrictions 3.

The structural signal matters more than the geopolitical one. Frontier-scale open weights used to function as a competitive moat: whoever shipped the biggest capable model shaped what developers could build with. Kimi K3 is the biggest. At 1.4 terabytes it is also, for everyone except the handful of operators who can keep that much memory warm, a benchmark flex. The open-weight movement has produced genuinely portable models at smaller scales, where a developer can pull weights onto a laptop and ship. At 2.8 trillion parameters, a permissive license changes ownership on paper while the hardware bill decides who can act on it. For most builders, the release that sounds like democratization is an invitation to rent.