Kimi K3, and what we can still learn from the pelican benchmark
Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.
2.8T-parameter Kimi K3 is now the largest open-weight model (weights by July 2026) and the first to hit $3/$15 per million input/output tokens. This sets a new floor for inference costs and GPU memory requirements—expect 16×A100s or 8×H100s just to load it, doubling your serving cluster size and cutting batch sizes in half, which will break any autoscaling rules tuned for 1T-class models.
Moonshot AI's Kimi K3 model costs $3 per million input tokens and $15 per million output tokens, making it one of the pricier models available, and generating a complex output like an SVG of a pelican riding a bicycle costs around 25 cents. This pricing will directly impact the cost of running production workloads for users shipping with this model, potentially increasing expenses. Running tasks that require large output tokens will be particularly costly.