Kimi-K2-Instruct

Checkpoint identity

Official identifier

moonshotai/Kimi-K2-InstructMoonshot AI Kimi-K2-Instruct model card (external source) · Kimi-K2-Instruct checkpoint configuration (external source) · Kimi K2 Modified MIT License (external source)

Architecture

Mixture-of-experts language model with Multi-head Latent AttentionMoonshot AI Kimi-K2-Instruct model card (external source) · Kimi-K2-Instruct checkpoint configuration (external source) · Kimi K2 Modified MIT License (external source)

Total parameters

1,000 billionPublished rounded counts; total stored weights determine the preliminary footprint, including all MoE experts.Moonshot AI Kimi-K2-Instruct model card (external source) · Kimi-K2-Instruct checkpoint configuration (external source) · Kimi K2 Modified MIT License (external source)

Active parameters

32 billion per tokenActive parameters are a compute-related count, not the model's full stored weight count.Moonshot AI Kimi-K2-Instruct model card (external source) · Kimi-K2-Instruct checkpoint configuration (external source) · Kimi K2 Modified MIT License (external source)

Modalities

text-input, text-outputMoonshot AI Kimi-K2-Instruct model card (external source) · Kimi-K2-Instruct checkpoint configuration (external source) · Kimi K2 Modified MIT License (external source)

Context

131,072 tokensNative configuration where directly verified. Extensions, advertised K-labels and configuration differences are preserved in the notes below.Moonshot AI Kimi-K2-Instruct model card (external source) · Kimi-K2-Instruct checkpoint configuration (external source) · Kimi K2 Modified MIT License (external source)

Release date

Exact release date not independently established in the reviewed artifactsMoonshot AI Kimi-K2-Instruct model card (external source) · Kimi-K2-Instruct checkpoint configuration (external source) · Kimi K2 Modified MIT License (external source)

Download access

Public checkpoint referenceMoonshot AI Kimi-K2-Instruct model card (external source) · Kimi-K2-Instruct checkpoint configuration (external source) · Kimi K2 Modified MIT License (external source)

License

Modified MITMIT-style permission and copyright notices have an added commercial attribution condition: products or services exceeding 100M monthly active users or US$20M monthly revenue must prominently display Kimi K2 in their interface. This is not plain MIT or an unconditional commercial approval.Moonshot AI Kimi-K2-Instruct model card (external source) · Kimi-K2-Instruct checkpoint configuration (external source) · Kimi K2 Modified MIT License (external source)

Theoretical weight-memory footprint

Using the publisher's approximately 1000 billion total parameters at 16 bits per parameter gives 2,000 GB (1862.6 GiB) of nominal weights.

This estimate excludes KV cache, activations, quantization metadata, non-weight buffers and serving overhead. It is not a tested GPU fit. MoE routing activates only part of the network per token while the complete expert weights still need storage.

Quantization

The selected checkpoint configuration uses blockwise FP8. Quantized derivatives and runtime kernels need independent compatibility and quality checks.

Serving considerations

• 1T total and 32B active are publisher-rounded approximate counts; the numeric total is 1,000 billion. All expert weights, cache and communication must be accounted for in deployment sizing.
• The card states 128K context and its configuration specifies 131,072 positions. This Instruct checkpoint is distinct from later Kimi K2 variants.
• Tool use needs Kimi-native parsing and correctly formatted tool messages. Aguasto has not evaluated or deployed the model; no facility or serving-capacity claim follows from this reference.

Framework documentation

These are third-party documentation starting points. Support depends on the exact checkpoint, runtime version and hardware.