Moonshot AI’s Kimi K3 hit the scene on July 17, pushing the open‑source frontier to 2.8 trillion parameters and a 1 million‑token context window — a scale previously seen only in proprietary labs. The YouTube demo showcases multimodal generation (text, images, audio) and a promotional 15 % discount for new users until September 1. While the model’s size is undeniable, the claim that it “beats GPT‑5.6” lacks third‑party verification; no benchmark suite (SWE‑bench, GPQA, MMLU) has been released, and the video provides only anecdotal prompts. For operators, the real question is whether the model’s inference cost and latency fit existing GPU clusters. Early community threads suggest that running Kimi 3 on a single RTX 5090 incurs roughly 30 W per token, which is higher than smaller open models but still cheaper than many closed‑source alternatives. Until formal performance data appear, buyers should treat the hype as a marketing hook and evaluate Kimi 3 against concrete workload metrics before committing hardware.
Sources
Composed by the MadCoolStuff editor pipeline · Groq · openai/gpt-oss-120b · 2026-07-17