Coding Horizon

Kimi K3 Has One Million Tokens And One Huge Trap

Every figure, price, date, threshold and benchmark this video puts on screen, chased to a primary source. The script arrived finished and was recorded before a shot existed, so anything here that could not be sourced is not drawn on screen at all, and anything the narration asserts without a primary source is listed at the bottom.

The model

On screen Value Source
Total parameters 2.8T moonshotai/Kimi-K3 model card
Activated parameters per token 104B moonshotai/Kimi-K3 model card
Number of experts 896 moonshotai/Kimi-K3 model card
Experts selected per token 16 moonshotai/Kimi-K3 model card
Context window 1,048,576 tokens moonshotai/Kimi-K3 model card
Vision encoder MoonViT-V2, 401M parameters, native text and image input moonshotai/Kimi-K3 model card
Quantisation MXFP4 weights, MXFP8 activations moonshotai/Kimi-K3 model card
Quantisation aware training applied from the SFT stage onward moonshotai/Kimi-K3 model card
Weights are published under a named licence Kimi K3 License moonshotai/Kimi-K3 model card

The model card is the primary source for all of the above: it is Moonshot’s own publication of the weights and the accompanying specification.

Benchmarks

Every number below is Moonshot’s own reported figure, from the same model card. The video says so on screen, because a score reported by the lab that trained the model is a different kind of claim from an independent one.

Benchmark Kimi K3 Source
DeepSWE 67.5 moonshotai/Kimi-K3 model card
Terminal-Bench 2.1 88.3 moonshotai/Kimi-K3 model card
FrontierSWE 81.2 moonshotai/Kimi-K3 model card
SWE-Marathon 42.0 moonshotai/Kimi-K3 model card
OmniDocBench 91.1 moonshotai/Kimi-K3 model card
MMMU-Pro 81.6 moonshotai/Kimi-K3 model card
Video-MME 90.0 moonshotai/Kimi-K3 model card
OSWorld-Verified 84.8 moonshotai/Kimi-K3 model card

The narration’s descriptions map onto those figures directly: DeepSWE “in the high sixties” is 67.5, Terminal Bench “around the high eighties” is 88.3, and FrontierSWE “above eighty in Moonshot’s table” is 81.2.

Price

On screen Value Source
Input, cache miss $3.00 per million tokens OpenRouter: moonshotai/kimi-k3
Input, cache hit $0.30 per million tokens OpenRouter: moonshotai/kimi-k3
Output $15.00 per million tokens OpenRouter: moonshotai/kimi-k3
Flat across the whole window, no long context surcharge OpenRouter: moonshotai/kimi-k3

The cache hit price is a 90% discount on the cache miss price, which is the arithmetic the pricing shot draws rather than a claim of its own.

The licence

On screen Condition Source
Model as a Service threshold a separate commercial agreement is required once the aggregate revenue of the licensee and its affiliates exceeds 20 million US dollars over any consecutive 12 month period Kimi K3 License analysis
Attribution threshold products above 100 million monthly active users, or 20 million US dollars in monthly revenue, must display the name Kimi K3 prominently in the interface Kimi K3 License analysis
Carve out: internal use activity that does not make the software, its outputs or its capabilities available to third parties Kimi K3 License analysis
Carve out: Moonshot’s own products use through Moonshot’s own offerings Kimi K3 License analysis
Carve out: certified inference partners approved third party providers Kimi K3 License analysis

The MaaS revenue gate counts the licensee’s revenue from all sources, not only revenue earned from Kimi K3. That is why the video draws it as a gate on the business rather than a gate on the model.

What is not put on screen

These are things the narration says that could not be chased to a primary source, so no shot renders a figure for them.