Kimi K3 Model Size: 2.8T Weights and API Cost

The viral Kimi K3 LLM City is a useful way to feel the size of a frontier open-weight model—but it is not a new Moonshot AI release or a price change. The independent WebGL project turns Kimi K3’s published architecture into a 3D “city” and a packed physical atlas.

Its headline count is 2,779,931,834,976 logical weight positions. That is consistent with Moonshot’s rounded official figure of 2.8 trillion total parameters. The live pricing table above still comes from our daily-maintained canonical dataset.

How large is Kimi K3 at 2.5 mm per weight?

LLM City lets the visitor choose an arbitrary physical edge length for each scalar. Its default is a 2.5 mm × 2.5 mm tile. Multiplying that area by the project’s logical count produces a striking result:

LLM City scale calculation	Result
Logical weight positions	2,779,931,834,976
Area per 2.5 mm square	0.00000625 m²
Total exact-area atlas	17,374,573.97 m²
Total in square kilometres	17.37 km²
Equal-area square	4.17 km × 4.17 km
FIFA-size football pitches	About 2,433

The physical scale is a metaphor, not a property of the checkpoint. Change the edge length and the area changes with its square. The valuable part is the relative geometry: the project maps tensor dimensions, expands the expert matrices, and exposes how different layers contribute to the total.

Does it really render every Kimi K3 weight?

The careful answer is: it represents the area and architecture of every logical weight position, but it does not publish the learned scalar values.

The project reports 249,924 logical tensor instances, 247,296 expert matrices, 497,220 physical checkpoint entries, and 12,068 symbolic forward-pass operations. It uses instanced geometry to make the architecture navigable in a browser. Its own detail panel says close-up cell numbers are deterministic mock values, not checkpoint contents. The creator made the same distinction in the Hacker News discussion: the values are not real, while the model layout is.

That boundary matters. LLM City is a strong architecture explainer and an auditable parameter-count model. It is not a new model-weight mirror, an inference engine, or evidence that trillions of individual values were independently drawn as separate browser objects.

Total parameters are not active parameters

Moonshot’s official model card calls Kimi K3 a 2.8T-parameter mixture-of-experts model with 104B activated parameters—about 3.7% of the total. The model selects 16 of 896 routed experts per token, alongside shared components.

Those figures answer different buyer questions:

• The 2.8T total helps explain checkpoint storage, accelerator memory, download time, and why self-hosting is an infrastructure project.
• The 104B active figure is closer to the sparse model capacity exercised for an individual token, though it still does not translate directly into latency or dollars.
• API cost depends on the provider’s token rate, caching, output length, latency, retries, and accepted-result rate—not the visualized land area.

Kimi K3’s official release uses MXFP4 weights and MXFP8 activations. In imec’s separate July self-hosting report, the roughly 1.4TB checkpoint did not leave enough KV-cache headroom on an 8×B200 node, so the test moved to 8×B300 with 2.3TB of HBM. That deployment evidence, not the tile metaphor, is the more practical starting point for hardware planning.

Did Kimi K3 API pricing change?

No. Moonshot’s official Kimi K3 rate card still lists the same cache-hit, cache-miss, and output categories already represented in our API pricing data. Telnyx and Novita remain separate managed routes with their own prices and serving terms.

Use the live table and calculator rather than copying a launch-day number into a spreadsheet. For variable demand, the managed API avoids a fixed multi-GPU bill. Self-hosting can make sense only when utilization, privacy, latency, control, and engineering requirements justify the infrastructure.

Teams that want Kimi K3 without operating a B300 node can compare Novita’s managed Kimi route with Moonshot and Telnyx.

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

Kimi K3 is now live in Labs

The complete August 17 AI Pricing Guru Labs run attempted all 49 deterministic text tasks through moonshotai/kimi-k3. It returned no endpoint errors and answered 47 correctly. The canonical direct-price estimate was $0.135252 for the run, while OpenRouter billed $0.13395165.

That result closes the earlier availability blocker. It does not test Kimi K3’s vision input, one-million-token context, self-hosted B300 throughput, or the accuracy of LLM City’s rendering. Labs measures cost per correct answer on a small deterministic public-endpoint suite.

Buyer takeaway

LLM City makes the storage scale of a 2.8T open-weight model memorable. It should not change a routing decision by itself.

API buyers should compare the live managed routes on total accepted-task cost. Self-hosters should size around the real checkpoint format, KV cache, concurrency, context, serving stack, and operational headroom. For a deeper route comparison, read our Kimi K3 API versus self-hosting analysis and the Moonshot AI pricing page.

Sources: the independent Kimi K3 LLM City visualization, the Hacker News discussion, Moonshot AI’s official Kimi K3 model card and weights, the official Kimi K3 API pricing page, imec’s self-hosting report, and the live AI Pricing Guru Labs dataset.