Kimi K3 matches GPT-5 on benchmarks, costs 4x less, and the weights are free

Kimi K3 requires approximately 1.56TB of storage in MXFP4 format, making local self-hosting extremely demanding. A community project called WASTE managed to run it on a 64GB MacBook Pro, but throughput was only 0.32–0.34 tokens per second due to SSD streaming bottlenecks — usable for experimentation but not production workloads. API access remains the practical path for most teams. Teams weighing self-hosted vs. API deployment for large open-weight models track real-world benchmarks like these on daily.dev.

Kimi K3 is available in GitHub Copilot across all plan tiers — Pro, Business, and Enterprise — hosted on Fireworks AI and billed at provider list pricing. Business and Enterprise admins must explicitly enable it. GitHub recommends reviewing open-weight model policies before enabling, signaling that compliance teams may need to weigh in before deployment. Developers navigating Copilot model options and enterprise AI policies find the latest changes covered on daily.dev.

Kimi K3 performs on par with or better than Claude and GPT-4o on most of 10 tested tasks spanning Blender and Godot — including shaders, physics simulations, particle effects, a Mario Kart clone, and a SimCity-style game. It achieves this at roughly 4x lower API cost than comparable closed models, with the added option of open weights for self-hosting. Engineers choosing between frontier models for game dev and creative coding tasks compare results like these on daily.dev.