Kimi K3 ships open-weight on Fireworks: 1,679 points tops the frontend code arena, training runs on a shared GPU pool
Synopsis
Open-weight Kimi K3 is available on Fireworks for inference and training from day one, which the release presents as frontier-level capability at a fraction of closed-model cost. On Arena's Frontend Code Arena it scores 1,679 points, ahead of Fable 5, making it the first open model to sit ahead of every closed one; it ranks third worldwide on the Artificial Analysis Intelligence Index at 57. Fireworks also offers US-only inference endpoints with zero data retention by default and a private preview of serverless training on a shared GPU pool.
Interpretation
Open-weight Kimi K3 can be run for inference and training on Fireworks the day it is released, positioned by the release as an open model that matches closed frontier quality on real work at far lower cost. The prior default was to start from a closed frontier model and route to specialized intelligence only where a task demanded it; the release says open models have crossed the line where they match closed frontier quality. The stated comparison is a 1,679-point score on Arena's Frontend Code Arena, ahead of Fable 5 and the first time an open model sits ahead of every closed one, plus third place worldwide on the Artificial Analysis Intelligence Index at 57.
K3's efficiency comes from changes to attention and mixture-of-experts structure rather than from size alone. Moonshot changed how information flows across the model, adding Kimi Delta Attention and refined attention residuals and scaling mixture of experts to activate 16 of 896 experts with Stable Latent MoE. The release reports 2.5 times the scaling efficiency of Kimi K2 for this combination, and Fireworks wrote custom kernels for Kimi Delta Attention, FP4 MoE and decode.
For regulated teams, Fireworks launches US-only serverless endpoints and keeps zero data retention by default for open models. Financial services, healthcare and other regulated companies requiring US data residency previously lacked this endpoint option. Kimi K3 is the first model on the US-only endpoint, with more popular models to follow over the coming days and weeks according to the release.
Perspective
The report covers what the Fireworks release states: K3's open-weight launch, day-one inference and training support, US-only serverless endpoints with zero data retention by default, and a private preview of serverless training. It addresses engineering and regulated teams deploying or fine-tuning frontier models.
The release states K3 matches Opus 5 at up to 5x better cost efficiency per task and that independent benchmarks such as Vals Index found the same, but these comparisons come from the release. Most leaderboard placements are third-party results relayed by the release, and evaluation configurations and sample details are not listed.