Can I self-host Kimi K3 on a single GPU?

No. Kimi K3 is a 2.8T-parameter Mixture-of-Experts model. Even in 4-bit quantization the weights exceed 1.4 TB, requiring a multi-GPU node such as 8×H100 80GB, 8×H200, or 8×MI300X. A single consumer GPU cannot load it in any usable form.

How is Kimi K3 priced through the official API?

$0.30 per 1M cached input tokens, $3 per 1M input tokens, and $15 per 1M output tokens. A top-up bonus of 10–30% runs until 11 August 2026.

When do the open weights land?

Moonshot announced the K3 open-weights release for 27 July 2026, roughly 11 days after the API launch on 16 July 2026.

At what monthly volume does self-hosting beat the API?

An 8×H100 on-demand node runs about $17,500/month. Breakeven against the API's $15 per 1M output rate is roughly 1.17 billion output tokens per month.

What about data residency and privacy?

If your OpenClaw workspace processes regulated data, self-hosting keeps prompts and completions inside your VPC. Otherwise the API is the right default.

Can I run a distilled K3 variant on cheaper hardware?

Distilled 32B–70B dense variants trained from K3 outputs will run on a single 80GB GPU or a pair of RTX 6000 Ada boards, and are the realistic self-host path for teams below the 1B token/month threshold.

Kimi K3 for OpenClaw: API vs Self-Hosted GPU (2026)

Kimi K3 is a 2.8T-parameter Mixture-of-Experts model with a 1M-token context window, released 16 July 2026 with open weights following on 27 July 2026. Kimi API pricing: $0.30 per 1M cached input, $3 per 1M input, $15 per 1M output.

Cost breakeven

An 8×H100 self-host node runs roughly $17,500/month at on-demand pricing. Breakeven against the API sits near 1.17 billion output tokens per month. Below that volume the API wins by a wide margin.

Hardware reality

K3 in INT4 quantization needs ~1.4 TB of VRAM — an 8×H100 80GB or 8×MI300X 192GB node. Consumer GPUs cannot serve the full model; distilled 32B–70B derivatives will be the realistic self-host path for smaller budgets.

Hybrid setup

Route reasoning and long-context requests to the Kimi K3 API; run a small quantized model locally for high-volume or private queries. OpenClaw supports per-model routing rules out of the box.