Kimi K3’s 1.4 TB weights fit only on 8×B300, not A100 or H200

qubridInc · reddit · 2026-07-27

Moonshot’s Kimi K3 weights are dropping, and one team is already doing the GPU math before the download lands.

The post says the model is expected to be about 2.8T total parameters, with 896 experts, 16 active experts per token, 1M context, and vision support. The weights are expected to be around 1.4 TB after MXFP4 quantization-aware training.

The author then works through deployment feasibility on three GPU stacks:

The model card is also described as unusually candid: quality drops if an agent harness truncates its thinking history, it tends to act instead of asking when ambiguous, and the chat experience is still said to lag models like Fable 5 and Sol even where benchmarks are close.

The team plans to publish throughput, time-to-first-token, and cost-per-million-token results across the three GPU configurations.

Related event: Moonshot Releases Open-Weight Kimi K3 Model(155 posts)→

Original post →

More from Infra

Infra channel →