Moonshot AI’s Kimi K3 has seized attention with 2.8 trillion parameters, making it the largest open-weight model announced to date. Yet its scale is not mainly a story about brute-force computation. The architecture repeatedly trades compute for memory, shifting the infrastructure bottleneck rather than removing it.
That distinction matters for organisations attracted by open weights, data sovereignty and the promise of lower serving costs.
K3 activates little but must keep everything ready
Kimi K3 uses a mixture-of-experts architecture with 896 specialised sections. Only 16—about 1.8% of the model—are activated for each token. This sharply reduces calculation compared with running all 2.8 trillion parameters every time.
The full model must still remain loaded because any expert may be selected next. Compute demand falls, but memory capacity and bandwidth remain substantial. Moonshot attacks that requirement through quantisation-aware training at four-bit precision, rather than a conventional 16-bit representation.
Independent analysis places the resulting weights near 1.4TB, compared with about 5.6TB at full precision. Kimi Delta Attention tackles the separate cache required when processing very long documents. Moonshot advertises a one-million-token context, where stored attention history can become larger than the model itself.
The company contributed caching code to the open-source vLLM serving project and recommends deploying K3 across at least 64 closely connected accelerators. Pooling memory across many chips resembles the logic of Huawei’s CloudMatrix systems. It may suit an environment where top-end training compute is restricted but large clusters can still be assembled.
Moonshot tested on Nvidia H200S, an unnamed alternative accelerator and an Nvidia L20 available under China-focused export restrictions. Its blog does not fully establish where the most capable hardware was located or how performance transfers to domestic silicon.
Open weights do not guarantee easy ownership
K3’s weights are scheduled for release on July 27, allowing organisations to download, modify and run the model internally. For banks, insurers and public institutions, that can support regional-language needs and keep sensitive records within a defined jurisdiction.
But a 1.4TB model spread across 64 accelerators is a data-centre project, not a server-room installation. Buyers must include high-bandwidth interconnects, spare capacity, power, cooling, monitoring and specialist engineering in the ownership calculation.
Most enterprises will therefore rent dedicated infrastructure. That can preserve contractual control and in-country processing without providing independence from a cloud or specialist host. The practical benefit is deployment choice, not free self-sufficiency.
Software readiness adds another constraint. K3’s architectural changes are new enough that common open-source serving tools do not yet support them fully. Moonshot is working with inference partners and maintainers, so the date weights become downloadable may precede the date a typical team can run them reliably.
API economics are not automatically budget-tier either. Pricing is $3 per million input tokens, $0.30 for cached input and $15 per million output tokens. Maximum reasoning effort is initially the only mode, meaning long reasoning chains and retries can dominate the cost of a completed task.
Performance claims need deployment-level verification
K3 ranked first in Arena’s blind Frontend Code evaluation at 1,679 points, ahead of a leading rival. That is useful evidence in one coding domain, not a universal verdict.
Moonshot itself says overall quality still trails Claude Fable 5 and GPT 5.6 Sol. It warns that output can become unstable if an application fails to return previous reasoning context, that the model may act unexpectedly when intent is ambiguous and that the user experience remains behind those competitors.
No organisation should choose K3 from parameter count alone. A serious evaluation should reproduce representative tasks, include memory and networking costs, measure completed-task latency and record how often tool calls or long-context sessions fail. Security, licensing and model-governance requirements must be assessed like any other self-hosted component.
Open-weight models already handled 29% of tokens through Vercel’s production gateway in June while accounting for under 4% of spending. K3 shows how that movement is entering a new infrastructure class. Its achievement is not escaping hardware limits; it is redesigning around them. Whether that trade works commercially will be decided by memory systems, serving software and operational discipline—not by the 2.8-trillion headline.
