Kimi K3 open weights drop claims best open model, but license isn't MIT

[AINews] Much ado about Open Weights

Moonshot AI shipped Kimi K3 over the weekend: a 2.8T total / 104B active MoE with 1M context and native vision. Artificial Analysis puts it ahead of Claude Opus 4.8 and only 4 points behind the top proprietary model. The tech report details MXFP4 weights, MXFP8 activations, and ~2.5× training efficiency gains, but total training tokens are missing. The license allows commercial use but requires a separate deal for cloud providers above $20M/year and a “Kimi K3” badge for products over 100M MAU or $20M/month. The same week saw an open-models letter signed by NVIDIA and Microsoft; OpenAI signed after rumors it wouldn't, Anthropic didn't. The author says skip the drama and read the tech report.

Why it matters: Moonshot's Kimi K3 open-weight release is the biggest domestic model launch this week — 2.8T params, beats Claude Opus 4.8 on benchmarks, only 4 points behind the top closed model. Not a 90+ because we only have the headline and summary so far; the full technical report detail...