Moonshot AI Kimi K3 Pricing

Managed API versus self-hosting

Moonshot's managed API is the simplest route for variable traffic: you pay for tokens and avoid owning an accelerator pool. Self-hosting exchanges that bill for fixed hardware capacity, serving engineering, monitoring, upgrades, and utilization risk.

The July 29 imec test makes the scale concrete. Its roughly 1.4TB Kimi K3 checkpoint did not fit on an 8×B200 node with enough room left for KV cache. The team moved to 8×B300, estimated the hardware at about 20% more than its GLM-5.2 B200 setup, and measured lower throughput and longer task time—but a 23.9-point task-resolution lead on its 64-task subset.

Read the complete methodology and contamination caveat in our Kimi K3 API versus self-hosting analysis. Use the token cost calculator to price managed volume before comparing it with a continuously rented B300 node.

Open weights do not mean zero cost

Moonshot describes Kimi K3 as a 2.8T-parameter mixture-of-experts model with 104B active parameters, MXFP4 weights, native multimodality, and a one-million-token context window. The weights are downloadable under the Kimi K3 License, but teams still need to review those terms and budget for storage, HBM, KV cache, power or rental, networking, and on-call operations.

If operating a multi-GPU node is not economical, compare Novita's managed Kimi K3 route with Moonshot's first-party API and other hosts.

Affiliate disclosure: we may earn a commission from the sponsored link above. It does not affect the pricing table or analysis.

What the Kimi K3 “LLM City” shows

The independent LLM City visualization maps Kimi K3's tensor shapes into a 3D architecture and an exact-area atlas. It counts 2,779,931,834,976 logical weight positions—consistent with Moonshot's rounded 2.8T total—and assigns every scalar a user-selectable physical tile. At the default 2.5 mm setting, that area is about 17.37 km², equal to a square roughly 4.17 km on each side.

This is an architecture-scale metaphor, not a dump of learned values. The page says its close-up cell numbers are deterministic mock values. Also keep total and active parameters separate: Moonshot says K3 activates about 104B parameters per token, so the 2.8T city is useful for understanding storage and deployment scale, not as a direct measure of per-token compute or API price.

Read our Kimi K3 model-size and LLM City cost analysis for the scale math and buyer implications.

What the four-SSD MacBook result proves

The independent ARGODRIVE Deltafin project reports 1.0015 tokens/s of steady decode for a 512-token raw completion on an M5 Max MacBook Pro with 128GB unified memory. The setup streamed Kimi K3's expert bank from the internal SSD and three external NVMe enclosures. The same package reports about 375 seconds to the first token for a 512-token prompt.

This is a reproducible systems experiment, not a Moonshot performance claim or a price change. Its four-drive ladder, project-specific storage layout, speculative drafter, int8 resident spine, cold-run method, and prompt dependence make the result unsuitable as a universal local-speed estimate. Read the Kimi K3 MacBook Pro cost and latency analysis for the measurement boundaries and managed-API comparison.

Labs status

Kimi K3 is now live in the AI Pricing Guru Labs leaderboard. The complete August 17 run attempted all 49 deterministic tasks with no endpoint errors and scored 47 correct. The direct-price estimate was $0.135252 for the run, while OpenRouter billed $0.13395165.

Labs still does not test the private imec task subset, SGLang configuration, B300 utilization, Deltafin's four-SSD local throughput, vision performance, or LLM City's architectural representation. It measures public-endpoint cost per correct answer on a small text suite. The local replay requirements are documented in our benchmark coverage notes.